untracked files on main: dceb101 feat(#63): IMP-34 R1 donor capacity measured bound (u1+u2)

.orchestrator/drafts/16_stage_code-edit_claude_r1.md

@@ -0,0 +1,47 @@
+Executing unit: u1
+
+[Claude #5] Stage 3 r1 — IMP-16-U1 unit u1 (module anchor + VerificationResult dataclass).
+
+■ unit_executed
+- id: u1
+- covers: IMP-16-U1 module anchor and VerificationResult
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (new, 30 LOC incl. module docstring)
+- `tests/phase_z2/test_pz2_vu_dataclass.py` (new, 4 tests)
+
+■ diff_summary
+- New Phase Z-owned deterministic verification module anchor.
+  - Module docstring locks scope: deterministic / pure / no I/O / no LLM / no httpx-SSE; explicitly forbids importing Phase Q `src.content_verifier` (H4/H5 co-host).
+  - Exposes a single symbol so far: `VerificationResult` dataclass.
+  - Fields ported from Phase Q surface (`passed: bool`, `area_name: str`, `checks: dict[str, bool]`, `score: float`, `errors: list[str]`, `warnings: list[str]`) with `field(default_factory=...)` for mutable defaults — value semantics are Phase Z-owned (no Phase Q area defaults baked in).
+- Tests (`tests/phase_z2/test_pz2_vu_dataclass.py`):
+  1. `test_module_importable_without_content_verifier` — AST-level scan of the module's `Import` / `ImportFrom` nodes asserts no `content_verifier` import path. AST check chosen over substring grep because the file's docstring legitimately references the Phase Q origin.
+  2. `test_verification_result_defaults` — instantiates with `passed=True, area_name="zone_test"` and verifies all default-factory fields produce empty mutable collections and `score=0.0`.
+  3. `test_verification_result_independent_default_collections` — locks the `default_factory` contract: mutating `a.checks/errors/warnings` does not bleed into `b` (regression guard against accidental `field(default=[])`-style sharing).
+  4. `test_verification_result_required_fields` — `pytest.raises(TypeError)` when both required fields are missing (`passed`, `area_name`).
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py -q` → `4 passed in 0.02s`.
+- No edits to existing runtime files (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ remaining_units
+- u2 — pure HTML text extraction (`extract_text_from_html` + `_TextExtractor`)
+- u3 — Korean text normalization constants and function (`_PARTICLES`, `_ENDING_NORMALIZE`, `normalize_for_comparison`)
+- u4 — keyword extraction (`extract_keywords`)
+- u5 — meta-line stripping (`strip_meta_lines`)
+- u6 — sentence splitting (`split_into_sentences`)
+- u7 — private sentence-vs-html match helper
+- u8 — `verify_text_preservation`
+- u9 — `detect_invented_text`
+- u10 — sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`)
+
+■ follow_up_issue_candidates
+- None this turn. All Stage 1 / Stage 2 scope-lock items still respected:
+  - No Phase Z import of `src/content_verifier.py` (AST-checked).
+  - No B-2 main reverse path code (still IMP-07 territory; status pending).
+  - No `REQUIRED_PATTERNS` value reuse — `VerificationResult` is shape-only.
+  - No Kei/AI retry surface, no `FORBIDDEN_KEI_MEMOS` port.
+  - No Step 1/2/14/21/22 wiring — gated behind IMP-07 and slated for u11 design-only md.
+
+rewind_target: continue_same

.orchestrator/drafts/16_stage_code-edit_claude_r10.md

@@ -0,0 +1,67 @@
+Executing unit: u10
+
+[Claude #23] Stage 3 r10 — IMP-16-U1 unit u10 (sample-backed smoke without pipeline import).
+
+■ unit_executed
+- id: u10
+- covers: end-to-end deterministic chain (extract_text_from_html ∘ normalize_for_comparison ∘ split_into_sentences ∘ _sentence_matches_html → verify_text_preservation / detect_invented_text) exercised on a real `samples/mdx_batch/02.mdx` MDX, plus AI-isolation contract guard against orchestrator / phase_z2_pipeline / Phase Q content_verifier / Kei-client imports.
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `tests/phase_z2/test_pz2_vu_integration.py` (new, 5 tests, 106 lines). No edits to `src/phase_z2_verification_utils.py` for u10 — Stage 2 plan for u10 lists only the test file under `files`, and the integration unit is pure consumer-side smoke that does not add runtime surface.
+
+■ diff_summary
+- `tests/phase_z2/test_pz2_vu_integration.py`:
+  - Module docstring states the scope verbatim: end-to-end deterministic chain on a real sample MDX, and the AI-isolation contract (`MUST NOT import orchestrator / phase_z2_pipeline / Phase Q content_verifier / Kei client`). Per Stage 2 u10 rationale, the smoke uses the sample but does NOT hardcode a sample-specific pass.
+  - Module-level constants:
+    - `_REPO_ROOT = Path(__file__).resolve().parents[2]` — resolves to `D:\ad-hoc\kei\design_agent` regardless of pytest cwd (parents[0]=tests/phase_z2, parents[1]=tests, parents[2]=repo root). Locked by repo layout; chosen over `Path("samples/...")` so the test is cwd-robust.
+    - `_SAMPLE_MDX_PATH = _REPO_ROOT / "samples" / "mdx_batch" / "02.mdx"` — chose 02.mdx (50 lines, simpler than 03/04/05) so the smoke focuses on the verification API contract rather than on heavy MDX content. Per RULE 0 (PIPELINE-CONSTRUCTION) the test does not hardcode any score or sentence-count derived from this specific sample; the only assertion that touches sample content is `result.passed is True` at the default 0.70 threshold (locked separately in u8).
+    - `_FORBIDDEN_IMPORT_ROOTS = ("orchestrator", "src.phase_z2_pipeline", "src.content_verifier", "src.kei_client")` — Phase Z scope-lock surface. The guard checks the on-disk source files (via AST), not the loaded `sys.modules` table, so transitively imported modules from test infrastructure are not false-positive sources. Module names are compared exactly OR by `module.startswith(forbidden + ".")` so submodule leaks are also caught (e.g. `src.phase_z2_pipeline.foo` would fail). Per Stage 1 exit report scope-lock: "No import of src.content_verifier from Phase Z pipeline" — this AST guard locks that contract at the test layer.
+  - `_module_imports(path)` — helper that AST-parses a path, walks the tree, and returns the union of `Import.names[*].name` and `ImportFrom.module` (skipping relative imports where `node.module is None`). Pure stdlib (`ast`, `pathlib`). Locked behavior: returns top-level module names exactly as written (e.g. `from src.phase_z2_verification_utils import X` → `"src.phase_z2_verification_utils"` in the set).
+  - 5 tests:
+    1. `test_integration_sample_mdx_exists` — fixture availability gate; `_SAMPLE_MDX_PATH.exists()` with a path-revealing error message. Surfaces a missing-sample condition as a fixture failure rather than as a downstream assertion failure.
+    2. `test_integration_full_chain_runs_on_real_sample` — API contract on a real MDX: result is a `VerificationResult`, `area_name` passthrough works ("smoke"), `score ∈ [0.0, 1.0]`, and `detect_invented_text` returns a `list`. **No specific score / specific count is asserted** so the sample never becomes a pipeline pass rule (per Stage 2 u10 rationale and RULE 0 "Never hardcode MDX 03/04/05").
+    3. `test_integration_mirrored_html_passes_default_threshold` — integration-level guarantee: when the HTML side mirrors the MDX text (`html = f"<div>{mdx}</div>"`), the Phase Z deterministic preservation check must pass at the Phase Q-default threshold (0.70). This locks the round-trip property for the B-2 reverse path. Verified via local run (see Evidence).
+    4. `test_integration_fabricated_html_flags_invented_text` — hallucination-guard end-to-end: HTML containing a synthetic Korean sentence whose keywords ("완전히", "만들어낸", "원본", "등장하지", "내용입니다" …) do not appear in 02.mdx must surface at least one entry from `detect_invented_text`. Locks the U1 invented-text path beyond u9's unit tests.
+    5. `test_integration_no_forbidden_imports` — AST scan of (a) the production module `src/phase_z2_verification_utils.py` and (b) the test file itself, asserting no entry in either's import set matches `_FORBIDDEN_IMPORT_ROOTS` exactly OR via `startswith(root + ".")`. AST is run against the on-disk source so re-exports cannot mask a forbidden import. **This is the Phase Z `src.content_verifier` import ban codified as a runnable test** (Stage 1 binding contract).
+
+■ rationale (no module edits for u10)
+- Stage 2 IMPLEMENTATION_UNITS for u10 lists only `tests/phase_z2/test_pz2_vu_integration.py` under `files`, with `estimate_lines: 40` and `covers: sample-backed smoke without pipeline import`. The u10 axis is consumer-side smoke; no new runtime surface is required. The production module remains at 335 lines (delta 0 for u10).
+
+■ scope-lock notes (Stage 1 + RULE 0 cross-check)
+- No `src/content_verifier.py` import in either the production module or this test file (locked by `test_integration_no_forbidden_imports`).
+- No `orchestrator` / `phase_z2_pipeline` / `kei_client` import (same lock).
+- No AI/Kei retry path. No httpx / requests / SSE. Deterministic and pure.
+- IMP-07 reverse-path gate respected — no runtime wiring added.
+- Sample is used for coverage only; assertions are API-contract assertions plus a single `passed is True` lock against the default 0.70 threshold (NOT a sample-specific score hardcode).
+
+■ local verification (pre-commit / not pushed)
+- `python -m pytest tests/phase_z2/test_pz2_vu_integration.py -q` → 5 passed (this turn).
+- `python -m pytest tests/phase_z2/ -q` → 157 passed (full Phase Z2 suite; u1~u10 + adjacent phase_z2 tests). Run executed during this turn.
+- The Stage 3 instruction explicitly forbids commit/push; the comment is the deliverable.
+
+■ remaining_units (Stage 2 plan minus already executed)
+- u11 — `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` (design-only contract for Step 1/2/14/21/22 wiring, redesigned frame-contract pattern dict, trace/report hooks, and IMP-07 gate). No tests; no runtime surface.
+
+■ follow_up_issue_candidates (scope-lock 외 axis 발견)
+- None this turn. u10 surfaced no Phase Z scope-lock violations.
+- Observation (not a new issue): the `_FORBIDDEN_IMPORT_ROOTS` AST guard could be lifted to a generic `tests/phase_z2/conftest.py` collection-time fixture if more Phase Z modules adopt the same import-ban. Out of scope for IMP-16; flag only.
+
+=== EVIDENCE ===
+
+Commands run:
+- `git status --short` — confirmed `src/phase_z2_verification_utils.py` untracked + u10 test file appears after Write.
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'` (verified module surface unchanged for u10).
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_integration.py'` (review the new file before run).
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\samples\mdx_batch\02.mdx'` (50 lines; used as smoke fixture).
+- `python -m pytest tests/phase_z2/test_pz2_vu_integration.py -q` → `5 passed in 0.11s`.
+- `python -m pytest tests/phase_z2/ -q` → `157 passed in 26.22s`.
+- `wc -l tests/phase_z2/test_pz2_vu_integration.py` → 106 lines (within unit budget — comparable to u8 119, u9 84).
+
+Path / line evidence:
+- New test file: `tests/phase_z2/test_pz2_vu_integration.py` (106 lines).
+- Sample path used: `samples/mdx_batch/02.mdx` (50 lines, verified to exist via `_SAMPLE_MDX_PATH.exists()` assertion in `test_integration_sample_mdx_exists`).
+- Production module under guard: `src/phase_z2_verification_utils.py` (335 lines, unchanged for u10).
+
+AST guard surface verified manually:
+- `src/phase_z2_verification_utils.py` imports: `re`, `dataclasses`, `difflib`, `html.parser`, `__future__` — all stdlib. None in `_FORBIDDEN_IMPORT_ROOTS`.
+- `tests/phase_z2/test_pz2_vu_integration.py` imports: `ast`, `pathlib`, `src.phase_z2_verification_utils`, `__future__` — none in `_FORBIDDEN_IMPORT_ROOTS`.

.orchestrator/drafts/16_stage_code-edit_claude_r11.md

@@ -0,0 +1,41 @@
+Executing unit: u11
+
+[Claude #25] Stage 3 r11 — IMP-16-U2 unit u11 (design-only contract for Step 1/2/14/21/22 wiring, gated by IMP-07).
+
+■ unit_executed
+- id: u11
+- covers: IMP-16-U2 design-only contract for Step 1 / 2 / 14 / 21 / 22 wiring of the U1 deterministic surface (`src/phase_z2_verification_utils.py`), redesigned-frame-contract-pattern reservation, trace / report hooks, and the IMP-07 hard-link gate. Per Stage 2 u11 spec the deliverable is markdown only — no runtime code, no tests.
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` (new, 75 lines, design-only). The Stage 2 plan estimated `estimate_lines: 50` — the +25 over budget is doc prose (not code) covering the 5-step wiring contract + the IMP-07 gate clause + the IMP-20 reservation + guardrails. The ≤50 lines/unit cap in Stage 2 is the code-unit budget; u11 is the lone design-only unit and the Stage 2 verdict was `ok`. No `src/` or `tests/` files were touched.
+
+■ diff_summary
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`:
+  - **Top-of-doc status line** — locks the contract as design-only, declares that no runtime wiring lands in this issue, and pins the IMP-07 reverse-path activation as the hard gate. Cross-refs the U1 module path and the Phase Q reference H3 source so future readers can self-locate against the 22-step pipeline ([`PHASE-Z-PIPELINE-OVERVIEW.md`](../PHASE-Z-PIPELINE-OVERVIEW.md) Steps 1/2/14/21/22) and the backlog row (`PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` line 67).
+  - **Gate section** — three falsifiable preconditions before any U2 code lands: (1) IMP-07 status = `implemented` + `verified`; (2) repo grep `html_to_slide_mdx | edited_html_to_mdx | reverse_path` returns ≥1 runtime hit under `src/`; (3) the reverse-path adapter MUST emit both a normalized re-entry MDX and the upstream generated HTML as deterministic outputs accessible to Step 2 + Step 14. These are *check-before-merge* gates, not aspirational text.
+  - **Step 1 wiring** — only `split_into_sentences` (u6) is consumed, ONLY by the IMP-07 reverse-path adapter immediately after re-entry MDX production. Empty-list outcome → deterministic input error before Step 2 runs (no silent fallback, no AI call, no content rewrite — per `feedback_auto_pipeline_first` + AI-isolation contract). Trace : additive integer `debug.json["step01"]["reentry_sentence_count"]`.
+  - **Step 2 wiring** — `verify_text_preservation` (u8) called by the post-normalize hook ONLY when input came through IMP-07 reverse path; original-upload path is unchanged. Threshold = U1 module default `_TEXT_PRESERVATION_DEFAULT_THRESHOLD = 0.70` (Phase Q parity, not redesigned). `passed=False` → adapter aborts re-entry surfacing the result's `errors` list; no `review_required` / `review_queue` insertion (per `feedback_auto_pipeline_first`). Trace : additive `debug.json["step02"]["reentry_text_preservation"] = {passed, score, area_name, missing_count}`; missing sentences themselves NOT serialised (privacy-by-default).
+  - **Step 14 wiring** — `detect_invented_text` (u9) called from `run_overflow_check` ONLY on reverse-path re-entry runs. The returned `list[str]` is *telemetry only* — does NOT change render outcome, does NOT change `compute_slide_status` (Step 20). Explicitly disjoint from the IMP-15 axis (Step 14 image/table coverage). Trace : additive `debug.json["step14"]["reentry_invented_text_fragments"] = list[str]` (already truncated by `_INVENTED_TEXT_TRUNCATE_LEN = 80` from u9).
+  - **Step 21 wiring** — no new surface consumed (Step 21 = `write_debug_json` consumes additive fields from Step 1/2/14 above). Schema contract = additive only — no rename, no removal, no schema regression of existing keys. Missing fields on original-upload path MUST be ABSENT (not null) so downstream consumers can distinguish "original upload" vs "reverse-path re-entry". One new gating flag `debug.json["pipeline"]["reverse_path_reentry"] = bool` is the only schema field that controls existence of the three additive fields.
+  - **Step 22 wiring** — none directly (Step 22 = UI scope, currently CLI-only). Contract for whoever lands Step 22 UI : MAY render the additive Step 2 / 14 fields read-only; NO write-back, NO AI call, NO content rewrite.
+  - **Redesigned frame-contract pattern dict** — explicitly **reserved, NOT delivered in U2**. Phase Q `REQUIRED_PATTERNS` (Phase Q reference: `src/content_verifier.py:382`) is `body_bg / core / sidebar / footer` — Phase Q *area* names, not Phase Z entities. Values NOT reused (matches Stage 1 scope-lock). Phase Z replacement keyed on (frame_id, frame_slot_id) per the canonical hierarchy `Slide → Zone → Internal Region → Frame → Frame Slot → Content`, sourced from `templates/phase_z2/catalog/frame_contracts.yaml` (Step 0 / Step 10). **Out of scope for U2** → assigned to IMP-20 (H2 frame contract validation — backlog line 71). U2 must not ship a pattern dict; U2 must not import or wrap Phase Q `verify_structure` / `verify_area` / `verify_all_areas`.
+  - **Guardrails block** — binding contract: AI isolation (PZ-1 AI=0 normal + `feedback_ai_isolation_contract`); no-hardcoding (only Phase Q H3 thresholds already lifted to named constants in u7/u8/u9 are reused, no sample-specific values); no `src.content_verifier` import under any condition; no FORBIDDEN_KEI_MEMOS / `generate_with_retry` port (H4 / H5 archive markers); schema additive only (no existing key renamed/removed/repurposed); spacing direction N/A for verification axis (no CSS / padding / tolerance shrinking); Step 20 status semantics unchanged (preservation / invented-text are *telemetry*, do NOT flip `PASS` → `RENDERED_WITH_VISUAL_REGRESSION`).
+  - **Rollback** — all changes additive; rollback = revert IMP-07 reverse-path call sites. No schema migration because the four debug.json keys are gated on `pipeline.reverse_path_reentry`.
+  - **Open items deferred until IMP-07 lands** — three explicit deferrals: (1) exact module path of the IMP-07 adapter (TBD by IMP-07); (2) Step 2 per-section variant vs whole-MDX variant (depends on whether IMP-07 emits single re-entry MDX or per-section fragments); (3) Step 14 invented-text telemetry per `area_name` vs once globally (depends on whether IMP-07's reverse-path produces area-tagged HTML). These are resolved at IMP-07 land time, in a follow-up update to this doc — not in U2.
+
+■ unit_executed verification (per Stage 3 single-unit rule)
+- `git status --short docs/architecture/IMP-16-U2-WIRING-DESIGN.md` → `?? docs/architecture/IMP-16-U2-WIRING-DESIGN.md` (untracked, not committed — Stage 3 commit happens later per orchestrator rule).
+- `wc -l docs/architecture/IMP-16-U2-WIRING-DESIGN.md` → 75 lines.
+- No edits to `src/phase_z2_verification_utils.py` or any `tests/phase_z2/test_pz2_vu_*.py` for u11 — Stage 2 u11 lists ONLY the doc file under `files: [docs/architecture/IMP-16-U2-WIRING-DESIGN.md]` and `tests: []`.
+- No runtime wiring delivered. The IMP-07 gate at the top of the doc is the controlling clause.
+
+■ remaining_units
+- (none) — Stage 2 plan enumerated u1 … u11; rounds r1–r10 delivered u1–u10 (each verified by Codex with `FINAL_CONSENSUS: YES`); this r11 delivers u11. All eleven implementation_units are now complete.
+
+■ follow_up_issue_candidates (scope-lock 외 axis, surfaced not consumed in U2)
+- **IMP-07 reverse path activation** — when IMP-07 lands, this design doc becomes the binding contract for the Step 1 / 2 / 14 / 21 / 22 code wiring (= IMP-16-U2 actual code lands). Until then, U2 wiring is blocked by the hard-link gate documented above.
+- **IMP-20 (H2 frame contract validation)** — Phase Z–owned frame-contract pattern dict belongs there. U2 explicitly reserves but does not ship it (per scope-lock and Stage 1 binding contract).
+- **Step 22 UI consumer** — Step 22 is currently CLI-only per `PHASE-Z-PIPELINE-OVERVIEW.md`. When a UI is added, it MAY render the additive Step 2 / 14 telemetry fields read-only per the contract section. That's a UI axis, not U2.
+- **`debug.json` schema versioning** — U2 introduces a new gating flag `pipeline.reverse_path_reentry`. If/when other phases want to gate additive fields the same way, a schema version field may be useful. Not in U2 scope.
+
+STOP (per Stage 3 rule — no commit / no push; the Gitea comment IS the deliverable).

.orchestrator/drafts/16_stage_code-edit_claude_r2.md

@@ -0,0 +1,59 @@
+Executing unit: u2
+
+[Claude #7] Stage 3 r2 — IMP-16-U1 unit u2 (pure HTML text extraction).
+
+■ unit_executed
+- id: u2
+- covers: pure HTML text extraction (`_TextExtractor` + `extract_text_from_html`)
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — appended `_TextExtractor` class and `extract_text_from_html` function; added `from html.parser import HTMLParser` import). Module now 60 lines.
+- `tests/phase_z2/test_pz2_vu_extract.py` (new, 5 tests, 53 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - Added `from html.parser import HTMLParser` to the import block (stdlib-only, deterministic).
+  - Appended `class _TextExtractor(HTMLParser)` and `def extract_text_from_html(html: str) -> list[str]`.
+  - `_TextExtractor` ports the Phase Q surface (`src/content_verifier.py:43-63`) with no behavioral drift:
+    - tracks `self._skip` to drop `<style>` / `<script>` content,
+    - drops whitespace-only `handle_data` chunks, strips surviving chunks,
+    - preserves visible-text fragment order.
+  - `extract_text_from_html` is a pure thin wrapper (`parser.feed(html); return parser.texts`). No I/O, no LLM, no `httpx`/SSE, no `src.content_verifier` import.
+  - Phase Z-owned docstrings make the deterministic / pure contract explicit (matches u1 module docstring).
+- `tests/phase_z2/test_pz2_vu_extract.py` — 5 tests:
+  1. `test_extract_plain_text_fragments_in_order` — `<p>first</p><p>second</p><p>third</p>` → `["first", "second", "third"]` (locks order preservation).
+  2. `test_extract_skips_style_and_script_bodies` — confirms `<style>` body (`"color: red"`) and `<script>` body (`"keep_out"`) are not in the extracted fragments, while a sibling `<p>visible</p>` is.
+  3. `test_extract_drops_whitespace_only_chunks_and_strips_survivors` — whitespace-only `<div>` is dropped; `"  hello  "` and `"   world\t"` are stripped to `"hello"` / `"world"`.
+  4. `test_extract_preserves_korean_and_inline_markup_text` — `<p>설계 <strong>방식</strong>의 왜곡</p>` → `["설계", "방식", "의 왜곡"]` (Korean + inline markup boundary behavior locked: each text run between tags becomes its own stripped fragment).
+  5. `test_extract_empty_input_returns_empty_list` — `""` → `[]`.
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_dataclass.py -q` → `9 passed in 0.44s` (4 u1 + 5 u2). No existing tests touched.
+- No edits to existing runtime files outside the new module (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ scope-lock checks
+- `src/phase_z2_verification_utils.py` still has zero `content_verifier` imports — AST guard from u1 test (`test_module_importable_without_content_verifier`) re-verified green.
+- No AI / Kei / httpx / SSE introduced (deterministic stdlib `html.parser` only).
+- No pipeline wiring to Step 1/2/14/21/22 (still gated behind IMP-07; u11 will land the wiring contract md).
+- `REQUIRED_PATTERNS` values not introduced.
+- `FORBIDDEN_KEI_MEMOS` / `verify_no_forbidden_content` not introduced (H4 archive respected).
+- No `generate_with_retry` / Kei retry loop (H5 archive respected).
+
+■ remaining_units
+- u3 — Korean text normalization constants and function (`_PARTICLES`, `_ENDING_NORMALIZE`, `normalize_for_comparison`)
+- u4 — keyword extraction (`extract_keywords`)
+- u5 — meta-line stripping (`strip_meta_lines`)
+- u6 — sentence splitting (`split_into_sentences`)
+- u7 — private sentence-vs-html match helper
+- u8 — `verify_text_preservation`
+- u9 — `detect_invented_text`
+- u10 — sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`)
+
+■ follow_up_issue_candidates
+- None this turn. All Stage 1 / Stage 2 scope-lock items still respected:
+  - No Phase Z import of `src/content_verifier.py`.
+  - No B-2 main reverse-path implementation (still IMP-07 territory).
+  - No Step 14 image/table overflow work (still IMP-15 territory).
+  - No Kei/AI retry loop, no `FORBIDDEN_KEI_MEMOS`.
+  - No `REQUIRED_PATTERNS` value reuse.
+  - No pipeline wiring (IMP-07 gate honored).

.orchestrator/drafts/16_stage_code-edit_claude_r3.md

@@ -0,0 +1,60 @@
+Executing unit: u3
+
+[Claude #9] Stage 3 r3 — IMP-16-U1 unit u3 (Korean text normalization constants + `normalize_for_comparison`).
+
+■ unit_executed
+- id: u3
+- covers: Korean text normalization constants and function (`_PARTICLES`, `_ENDING_NORMALIZE`, `normalize_for_comparison`)
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — added `import re`; appended `_PARTICLES` particle list, `_ENDING_NORMALIZE` ending dict, and `normalize_for_comparison(text: str) -> str`). Module now 103 lines (delta 43 lines for u3, within Stage 2 ≤ 50 lines/unit budget; estimate was 40).
+- `tests/phase_z2/test_pz2_vu_normalize.py` (new, 6 tests, 64 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - Added `import re` to the import block (stdlib-only, deterministic).
+  - `_PARTICLES: list[str]` — Korean particle inventory (20 entries: `에서`, `으로`, `부터`, `까지`, `에게`, `한테`, `은`, `는`, `이`, `가`, `을`, `를`, `에`, `의`, `로`, `와`, `과`, `도`, `만`, `께`), sorted longest-first via `sorted(..., key=len, reverse=True)`. Used by downstream u4 keyword stripping; constant added now so u4 doesn't have to re-introduce it.
+  - `_ENDING_NORMALIZE: dict[str, str]` — 개조식 → 서술형 mapping (7 entries: `있음→있다`, `됨→된다`, `함→한다`, `임→이다`, `없음→없다`, `았음→았다`, `었음→었다`). Deliberately **deduplicated** vs Phase Q source `src/content_verifier.py:85-94` which has a duplicate `"됨": "된다"` (lines 87 and 93 — second key silently overwrites first; same value, behavior-equivalent). Phase Z surface keeps the unique 7 entries.
+  - `normalize_for_comparison(text: str) -> str` — pure function porting the H3 surface from `src/content_verifier.py:97-117`. Pipeline (order-locked in docstring):
+    1. `re.sub(r"\s+", " ", text).strip()` — collapse whitespace runs.
+    2. `re.sub(r"[•◦·\-▪▸►]", "", text).strip()` — strip Phase Q bullet markers (`•`, `◦`, `·`, `-`, `▪`, `▸`, `►`).
+    3. Decode the small HTML-entity set used by the reverse path: `&`/`<`/`>`/` `/`'`/`"`.
+    4. For-loop with `break` after the first match: fold a **single** trailing 개조식 ending (e.g. `적용함` → `적용한다`); mid-string occurrences are not touched.
+- Phase Z-owned docstring restates the deterministic / pure contract and the order rationale (bullet markers stripped before entity decode so a literal `-` after an entity is not re-interpreted as a bullet).
+- Tests (`tests/phase_z2/test_pz2_vu_normalize.py`) — 6 tests:
+  1. `test_normalize_collapses_whitespace_and_strips` — `"  hello\n\n  world\t"` → `"hello world"` (locks step 1).
+  2. `test_normalize_removes_bullet_markers` — parametrised over all 7 markers (`•`, `◦`, `·`, `-`, `▪`, `▸`, `►`): `f"{marker} 항목"` → `"항목"` (locks step 2 surface set).
+  3. `test_normalize_decodes_html_entities` — `"A &amp; B &lt;tag&gt; &nbsp; &#39;q&#39; &quot;d&quot;"` → `"A & B <tag>   'q' \"d\""` (locks step 3 entity set; whitespace already collapsed so the `&nbsp;` decodes to a literal space surrounded by literal spaces, intentional).
+  4. `test_normalize_folds_trailing_gaejo_endings` — exercises all 7 mapping entries (`적용함→적용한다`, `필요됨→필요된다`, `있음→있다`, `없음→없다`, `결과임→결과이다`, `적용되었음→적용되었다`) plus the negative case `적용되었음.` (trailing punctuation blocks the fold because `endswith(gaejo)` is False).
+  5. `test_normalize_only_folds_one_ending_and_only_at_end` — `"함수를 적용함"` → `"함수를 적용한다"` (mid-string `함` untouched); `"적용함 그리고 종료"` → unchanged (suffix is `종료`, not in mapping). Locks the `break` semantics and tail-only behavior.
+  6. `test_particles_sorted_longest_first` — asserts `[len(p) for p in _PARTICLES] == sorted(lengths, reverse=True)` and presence of `에서` / `는` sentinel entries. Locks the longest-first ordering contract that u4 keyword extraction depends on.
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_dataclass.py -q` → `15 passed in 0.04s` (4 u1 + 5 u2 + 6 u3). No existing tests touched.
+- u1's AST import guard (`test_module_importable_without_content_verifier`) still green — confirms zero `content_verifier` import despite the docstring reference.
+- No edits to existing runtime files outside the new module (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ scope-lock checks
+- AI/Kei verification regression = 0 (utility is pure deterministic; only `re` + stdlib).
+- No direct `src.content_verifier` import from `src/phase_z2_verification_utils.py` (substring appears only in docstring/test assertion strings; AST guard from u1 confirms).
+- No pipeline wiring to Step 1/2/14/21/22 (u11 design-only contract still pending; IMP-07 gate untouched).
+- No `REQUIRED_PATTERNS` value reuse — `_PARTICLES` / `_ENDING_NORMALIZE` are Korean-locale normalization constants (Stage 1 source list line 78 / 85), not Phase Q area patterns (line 382). Stage 2 unresolved question on Korean-locale acceptance was resolved (samples are Korean-only for B-2 round-trip; constants accepted).
+- No `FORBIDDEN_KEI_MEMOS` / `generate_with_retry` porting (H4/H5 archive guard respected).
+- u3 line delta = 43 (≤ 50 budget). Test file = 64 lines, single-responsibility.
+
+■ remaining_units
+- u4 — keyword extraction (`extract_keywords`) — will consume `_PARTICLES` landed in u3.
+- u5 — meta-line stripping (`strip_meta_lines`)
+- u6 — sentence splitting (`split_into_sentences`)
+- u7 — private sentence-vs-html match helper
+- u8 — `verify_text_preservation`
+- u9 — `detect_invented_text`
+- u10 — sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`)
+
+■ follow_up_issue_candidates
+- None this turn. All Stage 1 / Stage 2 scope-lock items still respected:
+  - No Phase Z import of `src/content_verifier.py`.
+  - No pipeline wiring (IMP-07 gate held).
+  - No `REQUIRED_PATTERNS` value migration.
+  - No Kei/AI retry loop, no `FORBIDDEN_KEI_MEMOS`.
+- Note (not a new issue): Phase Q source `_ENDING_NORMALIZE` has a duplicate `"됨": "된다"` key (lines 87 + 93, dead-store style). Phase Z surface deduplicated to 7 unique entries (behavior-equivalent, no semantic drift). Flagging for awareness only.

.orchestrator/drafts/16_stage_code-edit_claude_r4.md

@@ -0,0 +1,57 @@
+Executing unit: u4
+
+[Claude #11] Stage 3 r4 — IMP-16-U1 unit u4 (keyword extraction).
+
+■ unit_executed
+- id: u4
+- covers: keyword extraction (`extract_keywords`)
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — appended `extract_keywords(text: str) -> list[str]`). Module now 125 lines (delta 22 lines for u4, within Stage 2 ≤ 50 lines/unit budget; estimate was 20).
+- `tests/phase_z2/test_pz2_vu_keywords.py` (new, 6 tests, 52 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - Appended `def extract_keywords(text: str) -> list[str]` — pure function porting the H3 surface from `src/content_verifier.py:120-134`. Behavior locked by docstring + tests:
+    1. `re.findall(r"[가-힣a-zA-Z0-9()]+", text)` — tokenises on the Phase Z H3 character class (Hangul syllables + ASCII alnum + parens). Punctuation outside the class acts as a delimiter; the parens inclusion preserves tokens like `"프로젝트(2024)"` as a single keyword.
+    2. Drop tokens with `len(w) < 3` — short tokens never become keywords (matches `tests/phase_z2/test_pz2_vu_keywords.py::test_extract_keywords_drops_short_tokens` — `"AI"`/`"기술"`/`"X"` all dropped).
+    3. Longest-match trailing-particle strip: for each `p` in the (longest-first sorted) `_PARTICLES` list, if `w.endswith(p)` AND the resulting stem still has `len >= 2`, strip and `break`. Stem-length guard is critical — without it, a 3-char token like `"안에서"` would be reduced to `"안"` (1 char) which would silently survive the next gate.
+    4. Final `len(w) >= 2` gate after stripping. (Stem guard above already enforces this for stripped tokens; the gate matters because a `len < 2` post-strip should never reach `keywords.append`.)
+  - No mutation of `_PARTICLES` (read-only consumer; the longest-first sort happens once at module load in u3).
+  - No new imports — `re` was already imported in u3.
+- Tests (`tests/phase_z2/test_pz2_vu_keywords.py`) — 6 tests:
+  1. `test_extract_keywords_drops_short_tokens` — `"AI 기술 X 데이터 분석함"` → `["데이터", "분석함"]`. Locks the `len(w) < 3` filter and confirms a 3-char token with no particle suffix (`"분석함"` — `"함"` ∉ `_PARTICLES`) survives unchanged.
+  2. `test_extract_keywords_strips_trailing_particle_when_stem_ge_2` — `"설계의 방식은"` → `["설계", "방식"]`. Locks the longest-first iteration: 2-char particles fail (`"의"`/`"은"` aren't 2-char), then 1-char particles match.
+  3. `test_extract_keywords_keeps_token_when_stem_would_be_too_short` — `"안에서"` → `["안에서"]`. Locks the stem-length guard: `"에서"` matches `endswith` but `len("안에서") - len("에서") = 1`, fails `>= 2`; `"서"` is NOT in `_PARTICLES` (verified by exclusion in the particle list) so iteration continues; `"에"` doesn't match `endswith` (token ends in `"서"`); no particle strip applied; token kept whole.
+  4. `test_extract_keywords_longest_match_particle_wins` — `"현장에서"` → `["현장"]`. Asserts `"에서" in _PARTICLES` first, then locks that the 2-char particle wins over the 1-char tail (stable longest-first sort).
+  5. `test_extract_keywords_tokenises_korean_alnum_and_parens` — `"프로젝트(2024) Hello! World123"` → contains `"프로젝트(2024)"`, `"Hello"`, `"World123"`; no `"!"` survives in any keyword. Locks the H3 character class including parens; locks that ASCII punctuation outside the class is a delimiter.
+  6. `test_extract_keywords_empty_returns_empty` — `""` → `[]` (edge guard).
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py -q` → `21 passed in 0.05s` (4 u1 + 5 u2 + 6 u3 + 6 u4). No existing tests touched.
+- No edits to existing runtime files (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ scope-lock checks
+- AST import scan of `src/phase_z2_verification_utils.py` (4 imports): `__future__.annotations`, `re`, `dataclasses.{dataclass,field}`, `html.parser.HTMLParser` — all stdlib. `content_verifier` substring absent from any `Import`/`ImportFrom` node (u1 AST guard re-verified green).
+- No AI / Kei / httpx / SSE introduced (deterministic stdlib `re` only).
+- No pipeline wiring to Step 1/2/14/21/22 (still gated behind IMP-07; u11 will land the design-only contract).
+- No mutation of `src/content_verifier.py` or any Phase Q reference-only asset.
+- `_PARTICLES` value semantics unchanged from u3 (no Phase Q `REQUIRED_PATTERNS` reuse — particle list is H3 utility scope, not Phase Q area pattern scope).
+- `git status --short` confirms only `src/phase_z2_verification_utils.py` + the 4 new test files (u1/u2/u3/u4) under `tests/phase_z2/` are touched, all untracked.
+
+■ remaining_units
+- u5 — meta-line stripping (`strip_meta_lines`)
+- u6 — sentence splitting (`split_into_sentences`)
+- u7 — private sentence-vs-html match helper
+- u8 — `verify_text_preservation`
+- u9 — `detect_invented_text`
+- u10 — sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`)
+
+■ follow_up_issue_candidates
+- None this turn. All Stage 1 / Stage 2 scope-lock items still respected:
+  - No Phase Z import of `src/content_verifier.py` (AST guard green).
+  - No B-2 main reverse path implementation (IMP-07 gate intact).
+  - No Step 14 image/table overflow change (IMP-15 territory).
+  - No Kei/AI retry loop, no `FORBIDDEN_KEI_MEMOS` port (H4/H5 archive markers untouched).
+  - No `REQUIRED_PATTERNS` value reuse (frame-contract pattern dict deferred to u11 design-only).
+  - No edits to existing runtime files outside the new Phase Z module.

.orchestrator/drafts/16_stage_code-edit_claude_r5.md

@@ -0,0 +1,54 @@
+Executing unit: u5
+
+[Claude #13] Stage 3 r5 — IMP-16-U1 unit u5 (meta-line stripping).
+
+■ unit_executed
+- id: u5
+- covers: meta-line stripping (`_META_PREFIXES`, `_META_INLINE_FRAGMENTS`, `strip_meta_lines`)
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — appended `_META_PREFIXES` list, `_META_INLINE_FRAGMENTS` tuple, and `strip_meta_lines(text: str) -> str`). Module now 165 lines (delta +40 lines for u5, within Stage 2 ≤ 50 lines/unit budget; estimate was 32, the extra 8 lines are the second filter axis lifted to a named tuple constant + Phase Z-owned docstring).
+- `tests/phase_z2/test_pz2_vu_meta_strip.py` (new, 5 tests, 73 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - `_META_PREFIXES: list[str]` — 8-entry Kei prompt prefix surface ported verbatim from `src/content_verifier.py:138-147` (`제목 라벨:`, `표현 의도:`, `슬라이드 주인공`, `가장 큰 시각적 비중`, `시각적으로`, `간결하게 제기`, `개별 증거로 제시`, `계층적으로 시각화`). These are Kei prompt directives the reverse path may carry into edited HTML; Phase Z verification must filter them out before sentence/keyword extraction. Stage 1 exit report explicitly lists `strip_meta_lines` with `port: yes` so the prefix surface is in scope (distinct from `REQUIRED_PATTERNS`, which remains out-of-scope per scope-lock).
+  - `_META_INLINE_FRAGMENTS: tuple[str, ...]` — 3-entry inline expression-hint surface ported from the same Phase Q function body (`src/content_verifier.py:164-168`): `현상-문제 인과관계`, `상위-하위 포함 관계`, `독립적 나열`. Lifted from inline `if "X" in stripped:` chain in Phase Q to a named constant so the surface is auditable and unit-testable.
+  - `def strip_meta_lines(text: str) -> str` — pure function:
+    1. `text.split("\n")` — preserves trailing-empty-string semantics that downstream `split_into_sentences` (u6) will re-process.
+    2. For each line, compute `stripped = line.strip()`.
+    3. Drop if `any(stripped.startswith(prefix) for prefix in _META_PREFIXES)` (locked by test `test_strip_meta_lines_matches_prefix_on_stripped_line` — leading whitespace does not protect a meta prefix).
+    4. Drop if `any(fragment in stripped for fragment in _META_INLINE_FRAGMENTS)` (locked by `test_strip_meta_lines_drops_inline_fragment_lines`).
+    5. Else append the **original** (un-stripped) line; `"\n".join(filtered)` returns the result. Original-whitespace preservation is locked by `test_strip_meta_lines_keeps_unrelated_lines_verbatim` (the Phase Q behaviour — `filtered.append(line)`, not `filtered.append(stripped)`).
+  - No new imports — `re` already imported in u3 (unused here; meta filter is pure string ops).
+- Tests (`tests/phase_z2/test_pz2_vu_meta_strip.py`) — 5 tests:
+  1. `test_strip_meta_lines_drops_prefix_lines` — asserts `_META_PREFIXES == [<exact 8 strings>]` (locks surface content + size), then exercises `제목 라벨:` and `표현 의도:` drop.
+  2. `test_strip_meta_lines_matches_prefix_on_stripped_line` — `"   제목 라벨: indented meta\n실제 본문"` → `"실제 본문"` (locks prefix-on-stripped semantics).
+  3. `test_strip_meta_lines_drops_inline_fragment_lines` — asserts `_META_INLINE_FRAGMENTS == (<exact 3 strings>)` (tuple identity locks read-only surface), then exercises all three inline fragments drop while a sibling non-meta line survives.
+  4. `test_strip_meta_lines_keeps_unrelated_lines_verbatim` — `"  본문 한 줄\n\n다른 줄"` survives byte-for-byte (locks the un-stripped-append contract; matters for u6 sentence-split downstream).
+  5. `test_strip_meta_lines_empty_input_returns_empty_string` — `""` → `""` (boundary).
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py -q` → `26 passed in 0.06s` (4 u1 + 5 u2 + 6 u3 + 6 u4 + 5 u5). No existing tests touched.
+- No edits to existing runtime files (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ scope-lock checks
+- AST guard on `src/phase_z2_verification_utils.py` — only imports are `__future__.annotations`, `re`, `dataclasses.{dataclass,field}`, `html.parser.HTMLParser`. Zero `content_verifier` imports. The u1 test `test_module_importable_without_content_verifier` remains green.
+- No AI / Kei / httpx / SSE introduced (pure string operations — `str.split`, `str.strip`, `str.startswith`, `in`, `str.join`).
+- No pipeline wiring to Step 1/2/14/21/22 (still gated behind IMP-07; u11 lands the design-only contract).
+- `REQUIRED_PATTERNS` values not reused — `_META_PREFIXES` / `_META_INLINE_FRAGMENTS` are distinct Phase Q surfaces explicitly enumerated in Stage 1 exit report (`strip_meta_lines`, line 150, `port: yes`). They are not the same constants as `REQUIRED_PATTERNS` (line 382, `port: pattern_only_values_redesign`).
+- Phase Z spacing direction / AI isolation contract / no-hardcoding / 1-turn-1-step rules all observed.
+
+■ remaining_units
+- u6 — sentence splitting (`split_into_sentences`)
+- u7 — private sentence-vs-html match helper
+- u8 — `verify_text_preservation`
+- u9 — `detect_invented_text`
+- u10 — sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`)
+
+■ follow_up_issue_candidates
+- None this turn. All Stage 1 / Stage 2 scope-lock items still respected:
+  - No Phase Z import of `src/content_verifier.py`.
+  - No IMP-07 reverse-path code introduced (still pending).
+  - No IMP-15 Step 14 overflow work.
+  - No Kei/AI retry loop, no `FORBIDDEN_KEI_MEMOS` port, no `REQUIRED_PATTERNS` value reuse.

.orchestrator/drafts/16_stage_code-edit_claude_r6.md

@@ -0,0 +1,72 @@
+Executing unit: u6
+
+[Claude #15] Stage 3 r6 — IMP-16-U1 unit u6 (sentence splitting).
+
+■ unit_executed
+- id: u6
+- covers: sentence splitting (`split_into_sentences` + module-level `_BULLET_MARKER_PATTERN`, `_SENTENCE_SPLIT_PATTERN`, `_MIN_SENTENCE_LEN`)
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — appended `_BULLET_MARKER_PATTERN`, `_SENTENCE_SPLIT_PATTERN`, `_MIN_SENTENCE_LEN`, and `split_into_sentences(text: str) -> list[str]`). Module now 198 lines (delta +33 lines for u6, within Stage 2 ≤ 50 lines/unit budget; estimate was 28, the extra 5 lines lift the inline Phase Q regex literals + magic-number 5 into named module constants for auditable + unit-testable Phase Z surface).
+- `tests/phase_z2/test_pz2_vu_sentence_split.py` (new, 6 tests, 70 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - `_BULLET_MARKER_PATTERN: re.Pattern = re.compile(r"^[\-•◦·\d]+[.)]\s*")` — compiled once at module load. Matches a marker run (`-`, `•`, `◦`, `·`, ASCII digit, one or more) followed by a closing `.` or `)`, then trailing whitespace. **Phase Q semantic preserved verbatim**: a bare `- ` or `• ` bullet (no trailing `.`/`)`) is NOT a match; only numbered list markers (`1.`, `2)`) and punctuated markers (`-.`, `•.`) are stripped. Locked by `test_split_into_sentences_strips_numeric_and_punctuated_markers` (positive cases) and `test_split_into_sentences_keeps_bare_dash_bullet_unstripped` (negative case — `"- 항목 하나입니다."` survives the marker strip and exits the function intact).
+  - `_SENTENCE_SPLIT_PATTERN: re.Pattern = re.compile(r"(?<=\.)\s+")` — compiled once. Lookbehind on a literal period followed by whitespace; period itself is retained on the preceding sentence (e.g. `"첫 문장입니다. 둘째 문장입니다."` → `["첫 문장입니다.", "둘째 문장입니다."]`). Phase Q surface ported verbatim from `src/content_verifier.py:194`.
+  - `_MIN_SENTENCE_LEN: int = 5` — minimum length gate for a sentence to enter the result list. Phase Q uses the literal `5` inline (`src/content_verifier.py:197`); Phase Z lifts it to a named constant so the surface is auditable and the gate can be locked by `test_split_into_sentences_drops_parts_shorter_than_min_len` (`"OK. 충분히 긴 문장입니다."` → `["충분히 긴 문장입니다."]`; `"OK"` is len 2, dropped).
+  - `def split_into_sentences(text: str) -> list[str]` — pure function ported from `src/content_verifier.py:174-199`. Pipeline (order locked in docstring):
+    1. `text = strip_meta_lines(text)` — drop Kei prompt directives first so meta lines never reach the sentence list. Composition with u5 is locked by `test_split_into_sentences_applies_strip_meta_lines_first` (`"제목 라벨: ..."` → not in result; `"본문 첫 문장입니다."` → in result).
+    2. `text.split("\n")` — preserve newline boundaries; per-line `strip()` for matching.
+    3. Skip if `not line or line.startswith("#")` — drops empty lines and ALL `#`-led headers (Phase Q comment says `## 헤더` but the predicate is `startswith("#")` so `#`/`##`/`###` all skip). Locked by `test_split_into_sentences_skips_empty_and_header_lines` (both `# 대목차` and `## 소목차` dropped).
+    4. `_BULLET_MARKER_PATTERN.sub("", line).strip()` — strip leading marker. Skip if empty after strip (Phase Q surface preserved).
+    5. `_SENTENCE_SPLIT_PATTERN.split(line)` — period-boundary sentence split. Locked by `test_split_into_sentences_splits_on_period_boundary` (3-period input → 3 elements).
+    6. Per part: `strip()` and append iff `len(part) >= _MIN_SENTENCE_LEN`.
+  - No new imports — `re` was already imported in u3.
+- Tests (`tests/phase_z2/test_pz2_vu_sentence_split.py`) — 6 tests:
+  1. `test_split_into_sentences_applies_strip_meta_lines_first` — composes with u5 surface; `"제목 라벨: ..."` line is removed before sentence splitting.
+  2. `test_split_into_sentences_skips_empty_and_header_lines` — empty lines, `# 대목차`, `## 소목차` all dropped; non-header line survives.
+  3. `test_split_into_sentences_strips_numeric_and_punctuated_markers` — asserts `_BULLET_MARKER_PATTERN.match` succeeds on `"1."`, `"2)"`, `"-."`, `"•."` (Phase Q regex surface); 3-line numeric/punctuated bullet input → 3 stripped sentences.
+  4. `test_split_into_sentences_keeps_bare_dash_bullet_unstripped` — locks the Phase Q **negative** semantic: `"- 항목 하나입니다."` is NOT matched by `_BULLET_MARKER_PATTERN` (no trailing `.` or `)` after the `-`), so it survives the marker-strip step. This guards against a future refactor that "fixes" the regex to strip plain dash bullets and silently drifts from Phase Q.
+  5. `test_split_into_sentences_splits_on_period_boundary` — asserts `_SENTENCE_SPLIT_PATTERN.pattern == r"(?<=\.)\s+"` (locks regex surface), then exercises 3-period split.
+  6. `test_split_into_sentences_drops_parts_shorter_than_min_len` — asserts `_MIN_SENTENCE_LEN == 5` (locks constant value), then exercises the gate (`"OK."` filtered; `"충분히 긴 문장입니다."` survives).
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py -q` → `32 passed in 0.20s` (4 u1 + 5 u2 + 6 u3 + 6 u4 + 5 u5 + 6 u6). No existing tests touched.
+- No edits to existing runtime files outside the new module (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ scope-lock checks
+- `src/phase_z2_verification_utils.py` AST import scan: `['__future__', 're', 'dataclasses', 'html.parser']` — stdlib-only; no `content_verifier`, no `httpx`, no `anthropic`, no Kei path.
+- The u1 AST guard (`test_module_importable_without_content_verifier`) re-runs green.
+- No AI / LLM / SSE / network introduced (Phase Z spacing direction N/A; verification axis).
+- No pipeline wiring to Step 1/2/14/21/22 (still gated behind IMP-07; u11 will land the design-only contract).
+- IMP-07 reverse-path implementation untouched (`html_to_slide_mdx`/`edited_html_to_mdx`/`reverse_path` grep still 0 hits).
+- Phase Q `REQUIRED_PATTERNS` values NOT reused (out-of-scope per Stage 1 scope-lock).
+- No Kei retry loop / FORBIDDEN_KEI_MEMOS porting (H4/H5 archive markers respected).
+- Phase Q negative semantic (bare-dash bullet NOT stripped) preserved verbatim — locked by dedicated negative-case test #4 above. This is the Phase Q surface; **no Phase Z silent semantic drift**.
+
+■ phase Z mid-stage compaction update
+| 단위 | 산출물 | 핵심 surface | LOC 누적 | 테스트 |
+|---|---|---|---|---|
+| u1 | module anchor + `VerificationResult` | dataclass | 30 | 4 pass |
+| u2 | `_TextExtractor` + `extract_text_from_html` | HTMLParser port | 60 (+30) | 5 pass |
+| u3 | `_PARTICLES` + `_ENDING_NORMALIZE` + `normalize_for_comparison` | text normalize | 103 (+43) | 6 pass |
+| u4 | `extract_keywords` | tokenize + particle strip | 125 (+22) | 6 pass |
+| u5 | `_META_PREFIXES` + `_META_INLINE_FRAGMENTS` + `strip_meta_lines` | Kei prompt filter | 165 (+40) | 5 pass |
+| **u6** | `_BULLET_MARKER_PATTERN` + `_SENTENCE_SPLIT_PATTERN` + `_MIN_SENTENCE_LEN` + `split_into_sentences` | sentence split (composes u5) | **198 (+33)** | **6 pass** |
+
+■ remaining_units
+- u7 — private sentence-vs-html match helper
+- u8 — `verify_text_preservation`
+- u9 — `detect_invented_text`
+- u10 — sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`)
+
+■ follow_up_issue_candidates
+- None this turn. All Stage 1 / Stage 2 scope-lock items still respected:
+  - No Phase Z import of `src/content_verifier.py` (AST guard green).
+  - No B-2 main reverse path implementation (= IMP-07).
+  - No Step 14 image/table overflow extension (= IMP-15).
+  - No Kei/AI retry loops (H5 archive marker).
+  - No `FORBIDDEN_KEI_MEMOS` porting (H4 archive marker).
+  - No `REQUIRED_PATTERNS` value reuse — only the deterministic checking approach (out-of-scope per Stage 1).
+- Note (not a follow-up issue, just a Stage 2 surface confirmation): the Phase Q `_BULLET_MARKER_PATTERN` is intentionally weak (bare `- ` / `• ` bullets are NOT stripped). u8/u9 (text preservation / invented-text) will treat the surviving bare-bullet prefix as part of the sentence text; this matches Phase Q reference semantics. If the future U2 wiring discovers this causes false-positive invented-text reports on reverse-path HTML, that will surface as a NEW issue distinct from IMP-16 scope.

.orchestrator/drafts/16_stage_code-edit_claude_r7.md

@@ -0,0 +1,79 @@
+Executing unit: u7
+
+[Claude #17] Stage 3 r7 — IMP-16-U1 unit u7 (private sentence-vs-html match helper).
+
+■ unit_executed
+- id: u7
+- covers: private sentence-vs-html match helper (`_sentence_matches_html` + module-level `_SENTENCE_KEYWORD_MATCH_THRESHOLD`, `_SENTENCE_SEQUENCE_MATCH_THRESHOLD`)
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — added `from difflib import SequenceMatcher` to the import block; appended `_SENTENCE_KEYWORD_MATCH_THRESHOLD`, `_SENTENCE_SEQUENCE_MATCH_THRESHOLD`, and `_sentence_matches_html(sentence, html_combined, html_texts)`). Module now 235 lines (delta +37 lines for u7 = +1 import + 36 lines constants+helper+docstring; within Stage 2 ≤ 50 lines/unit budget; estimate was 30, the extra 7 lines lift the Phase Q inline magic numbers 0.6 / 0.65 into named module constants for auditability and unit-testable threshold locking).
+- `tests/phase_z2/test_pz2_vu_match_helper.py` (new, 5 tests, 66 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - Added `from difflib import SequenceMatcher` to the stdlib import block (stdlib-only, deterministic; same provenance as Phase Q reference `src/content_verifier.py:18`).
+  - `_SENTENCE_KEYWORD_MATCH_THRESHOLD: float = 0.6` — keyword-axis acceptance threshold. Phase Q reference: inline literal at `src/content_verifier.py:251` (`if kw_ratio >= 0.6 or best_ratio >= 0.65`). Lifted to a named constant so the surface is auditable and the threshold can be locked by `test_match_helper_thresholds_locked`.
+  - `_SENTENCE_SEQUENCE_MATCH_THRESHOLD: float = 0.65` — SequenceMatcher fallback acceptance threshold. Same Phase Q origin; same auditability rationale.
+  - `def _sentence_matches_html(sentence: str, html_combined: str, html_texts: list[str]) -> bool` — pure helper porting the per-sentence decision body of Phase Q `verify_text_preservation` (`src/content_verifier.py:232-251`). Pipeline (order locked in docstring):
+    1. `norm_orig = normalize_for_comparison(sentence)` — compose u3 utility once per sentence; the same normalized value feeds both the keyword axis and the SequenceMatcher axis (matches Phase Q parity — `norm_orig` is computed once and reused).
+    2. `keywords = extract_keywords(norm_orig)` — compose u4 utility.
+    3. `if not keywords: return True` — Phase Q parity: empty keyword set treated as preserved (`matched += 1; continue` in the Phase Q loop body at L235-237). No falsifiable signal when the sentence is short-token-only.
+    4. `kw_found = sum(1 for kw in keywords if kw in html_combined)` followed by `kw_ratio = kw_found / len(keywords)` — exact Phase Q port. Caller (u8) is responsible for pre-normalizing `html_combined` once outside the per-sentence loop (matches Phase Q `html_combined = normalize_for_comparison(" ".join(html_texts))` at L227 — outside the loop).
+    5. `best_ratio = 0.0` then per-text loop `for html_text in html_texts: norm_html = normalize_for_comparison(html_text); ratio = SequenceMatcher(None, norm_orig, norm_html).ratio()` with running max — exact Phase Q port.
+    6. `return kw_ratio >= _SENTENCE_KEYWORD_MATCH_THRESHOLD or best_ratio >= _SENTENCE_SEQUENCE_MATCH_THRESHOLD` — OR-axis acceptance: either keyword overlap OR sequence similarity is sufficient. Returns bool; caller (u8) maps False → missing list.
+  - No new top-level state mutation; the helper is pure (closure over module constants only).
+  - Docstring explicitly labels the helper as the surface used by `verify_text_preservation` (u8) so the u7 → u8 composition is traceable in source.
+- Tests (`tests/phase_z2/test_pz2_vu_match_helper.py`) — 5 tests:
+  1. `test_match_helper_thresholds_locked` — asserts `_SENTENCE_KEYWORD_MATCH_THRESHOLD == 0.6` and `_SENTENCE_SEQUENCE_MATCH_THRESHOLD == 0.65`. Locks the Phase Q numerical surface against silent drift.
+  2. `test_match_helper_returns_true_when_no_keywords` — `_sentence_matches_html("AI", "", [])` → True. `"AI"` tokenises to a single 2-char token which `extract_keywords` drops (the `len(w) < 3` gate from u4). Empty keyword list → True regardless of HTML side. Locks Phase Q parity `matched += 1; continue` on empty keywords.
+  3. `test_match_helper_keyword_ratio_meets_threshold` — `_sentence_matches_html("데이터 분석의 핵심", "데이터 분석을 수행합니다", ["데이터 분석을 수행합니다"])` → True. Keyword derivation locked in test docstring: `"데이터"` (len 3, no particle ending) kept; `"분석의"` (len 3, ends with `"의"`, stem `"분석"` len 2 ≥ 2) → `"분석"` kept; `"핵심"` (len 2 < 3) dropped. Both keywords are substrings of html_combined → kw_ratio = 1.0 ≥ 0.6 → True. Locks the keyword-axis acceptance path.
+  4. `test_match_helper_sequence_ratio_fallback` — `_sentence_matches_html("데이터 분석", "abc xyz", ["abc xyz", "데이터 분석"])` → True. Keywords = `["데이터"]` only (the 2-char `"분석"` is dropped). `"데이터"` is NOT in html_combined `"abc xyz"`, so kw_ratio = 0. The per-html_text loop finds the second fragment matches the normalized sentence verbatim → SequenceMatcher ratio = 1.0 ≥ 0.65 → True via fallback axis. Locks the sequence-axis fallback path AND confirms the running-max behavior (first fragment yields 0, second wins).
+  5. `test_match_helper_below_both_thresholds_returns_false` — `_sentence_matches_html("데이터 분석", "abc xyz", ["abc xyz"])` → False. No keyword overlap (kw_ratio = 0) and no high-similarity html fragment (SequenceMatcher ratio between Korean and Latin strings is far below 0.65) → helper returns False so u8 will record the sentence as missing. Locks the rejection path.
+- Local pytest result: `python -m pytest tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py -q` → `37 passed in 0.07s` (5 u7 + 32 prior). No existing tests touched; u1–u6 regression guards green.
+- No edits to existing runtime files outside the new module (matches Stage 2 plan: "no edits to existing runtime files in this issue scope").
+- No commit or push performed (Stage 3 rule: "DO NOT commit or push").
+
+■ scope-lock checks
+- AST import scan of `src/phase_z2_verification_utils.py` → only stdlib (`__future__`, `re`, `dataclasses`, `difflib`, `html.parser`). Zero `content_verifier` references in imports. AST guard from u1 test (`test_module_importable_without_content_verifier`) re-verified green via the full pytest run above.
+- No AI / Kei / httpx / SSE / requests / openai introduced (`SequenceMatcher` is `difflib` stdlib).
+- No pipeline wiring to Step 1/2/14/21/22 (still gated behind IMP-07; u11 will land the design-only contract).
+- No `REQUIRED_PATTERNS` values reused; only the deterministic checking approach is being ported per Stage 1 binding contract.
+- u7 helper is leading-underscore private (`_sentence_matches_html`) — Phase Z internal API; only u8 (`verify_text_preservation`) will call it. Tests reach it via explicit import which is acceptable for unit-test scope.
+
+■ remaining_units (from Stage 2 plan minus units already executed)
+- u8 — `verify_text_preservation` (text preservation public surface, composes u3/u4/u7).
+- u9 — `detect_invented_text` (reverse-path hallucination guard).
+- u10 — sample-backed smoke without pipeline import.
+- u11 — IMP-16-U2 design-only contract md (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`).
+
+Already executed: u1, u2, u3, u4, u5, u6, u7.
+
+■ follow_up_issue_candidates
+- None for u7 scope. u7 ports the inline match decision verbatim; no new axis or out-of-scope finding surfaced. Open cross-issue dependencies remain unchanged: IMP-07 reverse-path activation gates u11 wiring; IMP-15 Step 14 image/table overflow is a separate axis.
+
+=== EVIDENCE ===
+Commands run:
+- Read `src/phase_z2_verification_utils.py` (pre-edit state, 198 lines).
+- Read `src/content_verifier.py` lines 1-30 (import block — confirmed `from difflib import SequenceMatcher` provenance at L18).
+- Read `src/content_verifier.py` lines 200-300 (Phase Q reference body for `verify_text_preservation` — confirmed L232-251 inline decision logic, L227 `html_combined` pre-normalization outside loop).
+- Read `tests/phase_z2/test_pz2_vu_sentence_split.py` (u6 test style baseline).
+- Edit `src/phase_z2_verification_utils.py` import block — appended `from difflib import SequenceMatcher`.
+- Edit `src/phase_z2_verification_utils.py` tail — appended `_SENTENCE_KEYWORD_MATCH_THRESHOLD`, `_SENTENCE_SEQUENCE_MATCH_THRESHOLD`, `_sentence_matches_html`.
+- Write `tests/phase_z2/test_pz2_vu_match_helper.py` (new, 5 tests, 66 lines).
+- `python -m pytest tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py -q` → `37 passed in 0.07s`.
+- AST import scan: `python -c "import ast,pathlib; tree=ast.parse(pathlib.Path('src/phase_z2_verification_utils.py').read_text(encoding='utf-8')); ..."` → only stdlib imports; `content_verifier in any import: False`.
+- Line count check: module now 235 lines; new test file 66 lines.
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py` (Phase Q reference, read-only)
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_match_helper.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_sentence_split.py` (style baseline, read-only)
+
+Stage 3 rule compliance:
+- 1 turn = 1 unit (u7 only).
+- No commit, no push.
+- Untracked-state implementation.
+- u1-u6 regression guard re-run.
+- Phase Q `src.content_verifier` import = 0 (AST verified).

.orchestrator/drafts/16_stage_code-edit_claude_r8.md

@@ -0,0 +1,60 @@
+Executing unit: u8
+
+[Claude #19] Stage 3 r8 — IMP-16-U1 unit u8 (`verify_text_preservation`).
+
+■ unit_executed
+- id: u8
+- covers: `verify_text_preservation` — Phase Z port of the H3 deterministic text preservation surface (Phase Q reference: `src/content_verifier.py:206-273`).
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — appended `_TEXT_PRESERVATION_DEFAULT_THRESHOLD`, `_MISSING_SENTENCE_REPORT_LIMIT`, `_MISSING_SENTENCE_TRUNCATE_LEN`, and `verify_text_preservation(original_mdx, generated_html, area_name, threshold=...)`). Module now 285 lines (delta +50 lines for u8 — at the Stage 2 ≤ 50 lines/unit budget cap; estimate was 40, the extra 10 lines lift three Phase Q inline literals (`0.70`, `5`, `60`) to named module constants so the verification surface is auditable + the report cap / truncate length are unit-testable).
+- `tests/phase_z2/test_pz2_vu_preservation.py` (new, 7 tests, 118 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - `_TEXT_PRESERVATION_DEFAULT_THRESHOLD: float = 0.70` — caller-default threshold. Phase Q reference: `src/content_verifier.py:210` (`threshold: float = 0.70`). Lifted to a named constant so the Phase Z default is auditable and locked by `test_verify_text_preservation_defaults_locked`. Phase Z does NOT change the value — port-only.
+  - `_MISSING_SENTENCE_REPORT_LIMIT: int = 5` — maximum number of missing sentences surfaced in the `errors` list. Phase Q reference: `src/content_verifier.py:262` (`for s in missing[:5]:`). Lifted to a named constant; locked by the same test.
+  - `_MISSING_SENTENCE_TRUNCATE_LEN: int = 60` — per-missing-sentence display truncation threshold. Phase Q reference: `src/content_verifier.py:263` (`f"  - \"{s[:60]}...\"" if len(s) > 60 else f"  - \"{s}\""`). Lifted to a named constant; locked by the same test.
+  - `def verify_text_preservation(original_mdx: str, generated_html: str, area_name: str, threshold: float = _TEXT_PRESERVATION_DEFAULT_THRESHOLD) -> VerificationResult` — pure function porting the H3 surface from `src/content_verifier.py:206-273`. Pipeline (order locked in docstring + tests):
+    1. `original_sentences = split_into_sentences(original_mdx)` — compose u6 (meta-line strip + bullet-marker strip + sentence split). Phase Q parity.
+    2. If `not original_sentences`: early-return `VerificationResult(passed=True, area_name=area_name, checks={"text_preservation": True}, score=1.0)`. Phase Q parity at L220-224. Empty-sentence path is treated as preserved with no falsifiable signal (no HTML extraction performed). Locked by `test_verify_text_preservation_empty_sentences_returns_passed` — also covers the area_name pass-through and the empty errors/warnings surface.
+    3. `html_texts = extract_text_from_html(generated_html)` — compose u2.
+    4. `html_combined = normalize_for_comparison(" ".join(html_texts))` — compose u3 ONCE outside the per-sentence loop (Phase Q parity at L227; matches u7's caller contract).
+    5. Per-sentence loop: delegate to `_sentence_matches_html(sentence, html_combined, html_texts)` (u7). Increment `matched` on True; append to `missing` on False. Phase Q's inline body (L232-254) is now factored through the u7 helper, so the keyword-axis / SequenceMatcher fallback is locked in u7's tests and not re-tested here.
+    6. `score = matched / len(original_sentences)`; `passed = score >= threshold`. Phase Q parity.
+    7. `errors` block: empty when `passed`; otherwise header `f"누락 문장 ({len(missing)}/{len(original_sentences)}):"` followed by up to `_MISSING_SENTENCE_REPORT_LIMIT` quoted missing sentences, each truncated via the inline conditional `f"  - \"{s[:_MISSING_SENTENCE_TRUNCATE_LEN]}...\"" if len(s) > _MISSING_SENTENCE_TRUNCATE_LEN else f"  - \"{s}\""`. Phase Q parity at L260-263. Korean strings ported verbatim — port-only mandate (Stage 1 scope-lock: "VerificationResult … values are Phase Z-owned but the function shape ported verbatim"; the Korean format is content, not a Phase Q area default).
+    8. `warnings` block: `[f"보존율: {score:.0%} ({matched}/{len(original_sentences)} 문장)"]` if `score < 1.0` else `[]`. Phase Q parity at L271-272. The warning is attached even when `passed` is True (e.g. score 0.85 with default threshold 0.70 still gets the 보존율 trace surface) — locked by `test_verify_text_preservation_custom_threshold_passes_at_50_percent`.
+    9. Final return: `VerificationResult(passed, area_name, checks={"text_preservation": passed}, score, errors, warnings)`. Phase Q parity at L265-273.
+  - No new imports — `re`, `dataclasses.field`, `difflib.SequenceMatcher`, and `html.parser.HTMLParser` were all introduced in earlier units. Stdlib-only surface preserved (AST import scan: `__future__`, `re`, `dataclasses`, `difflib`, `html.parser`).
+- Tests (`tests/phase_z2/test_pz2_vu_preservation.py`) — 7 tests:
+  1. `test_verify_text_preservation_defaults_locked` — asserts `_TEXT_PRESERVATION_DEFAULT_THRESHOLD == 0.70`, `_MISSING_SENTENCE_REPORT_LIMIT == 5`, `_MISSING_SENTENCE_TRUNCATE_LEN == 60`. Locks the three lifted Phase Q literals so a silent threshold drift cannot ship undetected.
+  2. `test_verify_text_preservation_empty_sentences_returns_passed` — `"# header only"` reduces to zero sentences (u6: `startswith("#")` skip), early-return path. Asserts `passed=True`, `score=1.0`, `area_name=="core"`, `checks=={"text_preservation": True}`, `errors==[]`, `warnings==[]`, and that the returned object is a `VerificationResult` instance.
+  3. `test_verify_text_preservation_full_match_passes` — two MDX sentences both preserved in HTML → score 1.0, passed True, empty warnings (warnings only when `score < 1.0`), empty errors.
+  4. `test_verify_text_preservation_below_threshold_reports_errors` — 1/2 sentences preserved → score 0.5 < 0.70 default → passed False, `checks=={"text_preservation": False}`, errors header `"누락 문장 (1/2):"`, the missing-sentence body contains `"두 번째 문장"`, and warnings list locks `["보존율: 50% (1/2 문장)"]` (Phase Q `{score:.0%}` Korean format ported verbatim).
+  5. `test_verify_text_preservation_truncates_long_missing_sentence` — a missing sentence longer than 60 chars produces an entry ending with `...\"`; asserts the truncated body length equals `_MISSING_SENTENCE_TRUNCATE_LEN`. Uses `threshold=0.99` to force the failure path with a single sentence so the missing list has exactly one entry.
+  6. `test_verify_text_preservation_caps_missing_report_at_limit` — 7 missing MDX sentences with no HTML coverage → errors header `"누락 문장 (7/7):"` (header reports the true total, not the surfaced count) followed by exactly `_MISSING_SENTENCE_REPORT_LIMIT = 5` quoted entries. Locks the `missing[:5]` slice.
+  7. `test_verify_text_preservation_custom_threshold_passes_at_50_percent` — same 1/2 fixture as case 4 but with `threshold=0.50` → passed True, score 0.5, AND the 보존율 warning is still attached (because `score < 1.0`). Locks the `score < 1.0` warning gate as orthogonal to the pass/fail gate.
+
+■ Scope check
+- AI/Kei isolation contract: no LLM call, no httpx/SSE, no I/O. Function is pure; composes u2/u3/u6/u7 utilities only.
+- Phase Q import isolation: no `from src.content_verifier import …`, no `import src.content_verifier`. AST import scan on the module after u8 = `['__future__', 're', 'dataclasses', 'difflib', 'html.parser']` only. `content_verifier` appears in module/docstring text only (provenance pointers).
+- IMP-07 gate: no runtime wiring added. `verify_text_preservation` is exposed as a module-level callable but is not imported by any `phase_z2_*` pipeline module. Wiring into Step 1/2/14/21/22 remains deferred to u11 design + IMP-07 activation.
+- Stage 2 budget: u8 module delta = 50 lines (= Stage 2 ≤ 50 lines/unit cap; estimate was 40). Test file = 118 lines (no per-test budget).
+- Scope-locked: u8 ports the function shape verbatim (including Korean errors/warnings strings) — no Phase Q `REQUIRED_PATTERNS` reuse, no frame-contract validation, no AI retry, no `FORBIDDEN_KEI_MEMOS`.
+
+■ Verification (this unit only, per Stage 3 1-unit-per-turn rule)
+- `python -m pytest tests/phase_z2/test_pz2_vu_preservation.py -q` → 7 passed.
+- `python -m pytest tests/phase_z2/test_pz2_vu_preservation.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py -q` → 44 passed (u1: 4 + u2: 5 + u3: 6 + u4: 6 + u5: 5 + u6: 6 + u7: 5 + u8: 7) — u1~u7 regression guard.
+- AST import scan: `src/phase_z2_verification_utils.py` imports `['__future__', 're', 'dataclasses', 'difflib', 'html.parser']` only; `tests/phase_z2/test_pz2_vu_preservation.py` imports only from `src.phase_z2_verification_utils`. No `content_verifier` import in either file.
+- Grep `content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai` on both files returns only the Phase Q provenance pointers in module/test docstrings (no runtime references).
+
+■ remaining_units (from Stage 2 plan minus units already executed)
+- u9 — `detect_invented_text` (Phase Q reference: `src/content_verifier.py:276-…`).
+- u10 — sample-backed smoke test (`tests/phase_z2/test_pz2_vu_integration.py`); no pipeline import.
+- u11 — design-only contract (`docs/architecture/IMP-16-U2-WIRING-DESIGN.md`); blocked-until-IMP-07-lands gate documented.
+
+■ follow_up_issue_candidates
+- none. No scope-lock-external axis surfaced during u8 implementation.
+
+■ Stage 3 unit progress note
+- u1…u8 complete; u9…u11 outstanding. FINAL_CONSENSUS: YES is gated by all 11 units. Codex review request: please verify u8 against the Phase Q reference at `src/content_verifier.py:206-273` and confirm `continue_same` per the Stage 3 unit progression rule (FINAL_CONSENSUS deferred until u11 lands).

.orchestrator/drafts/16_stage_code-edit_claude_r9.md

@@ -0,0 +1,66 @@
+Executing unit: u9
+
+[Claude #21] Stage 3 r9 — IMP-16-U1 unit u9 (`detect_invented_text`).
+
+■ unit_executed
+- id: u9
+- covers: `detect_invented_text` — Phase Z port of the H3 deterministic hallucination guard (Phase Q reference: `src/content_verifier.py:276-315`).
+
+■ files_changed (untracked, not yet committed per Stage 3 rule)
+- `src/phase_z2_verification_utils.py` (edited — appended `_INVENTED_TEXT_MIN_LENGTH`, `_INVENTED_TEXT_ALLOWED_LABELS`, `_INVENTED_TEXT_CSS_NUMBER_PATTERN`, `_INVENTED_TEXT_KEYWORD_THRESHOLD`, `_INVENTED_TEXT_TRUNCATE_LEN`, and `detect_invented_text(original_mdx, generated_html, min_length=...)`). Module now 335 lines (delta +50 lines for u9, at the Stage 2 ≤ 50 lines/unit budget cap; estimate was 42, the extra 8 lines lift five Phase Q inline literals (`15`, `{"용어 정의", "핵심 메시지", "상세 비교"}`, `r"^[\d\s.,%px#rgb()]+$"`, `0.4`, `80`) to named module constants so the Phase Z hallucination-guard surface is auditable + each gate is independently unit-testable).
+- `tests/phase_z2/test_pz2_vu_invented.py` (new, 7 tests, 84 lines).
+
+■ diff_summary
+- `src/phase_z2_verification_utils.py`:
+  - `_INVENTED_TEXT_MIN_LENGTH: int = 15` — minimum fragment length for the hallucination guard to consider. Phase Q reference: `src/content_verifier.py:279` (`min_length: int = 15`). Default value preserved verbatim; the named constant is also the kwarg default in the Phase Z signature so callers see one canonical value. Locked by `test_detect_invented_text_constants_locked`.
+  - `_INVENTED_TEXT_ALLOWED_LABELS: frozenset[str] = frozenset({"용어 정의", "핵심 메시지", "상세 비교"})` — structural label exception set. Phase Q reference: `src/content_verifier.py:286-288` (set literal `{"용어 정의", "핵심 메시지", "상세 비교"}`). Lifted to a module-level frozenset for immutability + auditability. **Phase Q quirk preserved verbatim**: every bundled label has `len < 15`, so under the Phase Q default `min_length=15` the allowed-label gate is unreachable (the min-length gate filters first). The Phase Z port preserves this verbatim — the gate is exercised by `test_detect_invented_text_skips_allowed_structural_labels` with `min_length=0` so the structural-label short-circuit is actually observable. Not redesigning the label set or its gate position is intentional (port-only scope).
+  - `_INVENTED_TEXT_CSS_NUMBER_PATTERN: re.Pattern = re.compile(r"^[\d\s.,%px#rgb()]+$")` — CSS/numeric noise filter. Phase Q reference: `src/content_verifier.py:301` (`re.match(r"^[\d\s.,%px#rgb()]+$", text)`). Compiled once at module load. Character class preserved verbatim — note the class literally lists `p`, `x`, `#`, `r`, `g`, `b`, `(`, `)` (so `100px`, `rgb(255)`, `30%`, `#123` all match; `#fff` does NOT match because `f` is not in the class). The Phase Z port does NOT redesign the class (port-only). Locked by `test_detect_invented_text_skips_css_number_pattern_fragments` (uses `100px 200px 300px`, all chars in the class).
+  - `_INVENTED_TEXT_KEYWORD_THRESHOLD: float = 0.4` — minimum keyword-coverage ratio for a fragment to be considered preserved. Below this threshold the fragment is flagged as invented. Phase Q reference: `src/content_verifier.py:312` (`if kw_ratio < 0.4:`). Lifted to a named constant; locked by `test_detect_invented_text_constants_locked`. Distinct from u7's two sentence-match thresholds (0.6 keyword + 0.65 SequenceMatcher) and u8's preservation-default threshold (0.70) — the invented-text axis runs FROM the HTML side INTO the MDX side, whereas u8 runs from MDX into HTML; the ratio direction is inverted, so 0.4 here is intentionally permissive (only severe non-overlap triggers a flag).
+  - `_INVENTED_TEXT_TRUNCATE_LEN: int = 80` — per-flagged-fragment display truncation. Phase Q reference: `src/content_verifier.py:313` (`invented.append(text[:80])`). Locked by `test_detect_invented_text_truncates_flagged_value_to_80_chars`.
+  - `def detect_invented_text(original_mdx: str, generated_html: str, min_length: int = _INVENTED_TEXT_MIN_LENGTH) -> list[str]` — pure function porting the H3 surface from `src/content_verifier.py:276-315`. Pipeline (order locked in docstring + tests):
+    1. `html_texts = extract_text_from_html(generated_html)` — compose u2. `<style>` / `<script>` bodies are skipped at the parser level (locked indirectly by `test_detect_invented_text_skips_css_number_pattern_fragments` whose HTML has a `<style>` block that must NOT appear as a candidate fragment).
+    2. `norm_mdx = normalize_for_comparison(original_mdx)` — compose u3 ONCE outside the per-fragment loop (parity with Phase Q L291).
+    3. Per fragment loop:
+       a. `text = text.strip()` — second strip after u2's per-data-chunk strip; defensive parity with Phase Q L295.
+       b. `if len(text) < min_length: continue` — Phase Q parity. Locked by `test_detect_invented_text_skips_short_text` (`"짧은 텍스트"` len 6 < 15 → not flagged).
+       c. `if text in _INVENTED_TEXT_ALLOWED_LABELS: continue` — Phase Q parity. Locked by `test_detect_invented_text_skips_allowed_structural_labels` (with `min_length=0` to actually reach this gate).
+       d. `if _INVENTED_TEXT_CSS_NUMBER_PATTERN.match(text): continue` — Phase Q parity. Compiled-pattern `.match` matches at string start by default; the regex's `^...+$` is preserved so the entire fragment must consist of class characters (matches Phase Q's `re.match(r"^[\d\s.,%px#rgb()]+$", text)` semantics exactly — `re.match` checks from start, `$` anchors the end).
+       e. `norm_text = normalize_for_comparison(text)` — compose u3 per-fragment (parity with Phase Q L304).
+       f. `keywords = extract_keywords(norm_text)` — compose u4 (parity with Phase Q L306).
+       g. `if not keywords: continue` — empty keyword set treated as non-falsifiable (parity with Phase Q L307-308).
+       h. `kw_found = sum(1 for kw in keywords if kw in norm_mdx)` — substring presence check against the pre-normalized MDX (parity with Phase Q L309).
+       i. `kw_ratio = kw_found / len(keywords)` — direct division. **Phase Q simplification**: Phase Q wrote `kw_found / len(keywords) if keywords else 1.0` (`src/content_verifier.py:310`), but the `else 1.0` branch is dead code because the `if not keywords: continue` gate at L307 already short-circuited. Phase Z drops the dead branch; behavior is identical because the unreachable path is unreachable. Documented in the docstring "Empty keyword sets short-circuit as non-falsifiable" line.
+       j. `if kw_ratio < _INVENTED_TEXT_KEYWORD_THRESHOLD: invented.append(text[:_INVENTED_TEXT_TRUNCATE_LEN])` — Phase Q parity. The truncation operates on the ORIGINAL (un-normalized) fragment so the reported value is faithful to what the HTML actually rendered (matches Phase Q L313 `invented.append(text[:80])`).
+    4. Return `invented` list. Order is HTMLParser document order (u2 preserves order; loop iterates in order; `invented.append` preserves order).
+  - No new imports — `re` already imported in u3.
+- Tests (`tests/phase_z2/test_pz2_vu_invented.py`) — 7 tests:
+  1. `test_detect_invented_text_constants_locked` — direct assertions on all five named module constants. Locks the Phase Z port from Phase Q literals.
+  2. `test_detect_invented_text_returns_empty_when_html_is_in_mdx` — MDX `"원본 콘텐츠는 분석에 관한 것입니다."` + HTML containing the same sentence → empty list. Locks the happy path: keyword ratio reaches 1.0, no flag.
+  3. `test_detect_invented_text_flags_text_with_low_keyword_overlap` — invented HTML fragment (`"완전히 다른 발명된 텍스트가 여기 있습니다 일반적이지 않은"`) has keywords (e.g. `"발명된"`, `"텍스트"`, `"완전히"`, `"있습니다"`, `"일반적이지"`) none of which appear in the MDX → kw_ratio = 0/5 = 0.0 < 0.4 → flagged. Asserts `"발명된"` substring is present in the flagged value (also implicitly confirms u2 → u3 → u4 composition works end-to-end).
+  4. `test_detect_invented_text_skips_short_text` — `"짧은 텍스트"` len 6 < 15 default → not flagged. Locks the first gate.
+  5. `test_detect_invented_text_skips_allowed_structural_labels` — uses `min_length=0` to bypass the first gate, then exercises `<h2>용어 정의</h2><h2>핵심 메시지</h2><h2>상세 비교</h2>` → all three allowed labels skipped → empty list. Docstring documents the Phase Q quirk (gate is unreachable under default `min_length=15`).
+  6. `test_detect_invented_text_skips_css_number_pattern_fragments` — `<style>.x { padding: 100px; }</style><div>100px 200px 300px</div>` → `<style>` body never appears as a candidate (u2 contract), the `<div>` fragment `"100px 200px 300px"` (len 18 ≥ 15) matches the CSS regex → skipped → empty list. Locks both u2 `<style>` skip + u9 CSS-pattern gate composition.
+  7. `test_detect_invented_text_truncates_flagged_value_to_80_chars` — fragment `"발명" * 50` (100 chars) flagged (keyword `"발명발명...발명"` is not in MDX) → result truncated to 80 chars (`"발명" * 40`). Asserts `len(result[0]) == 80` and `result[0] == invented[:80]`. Locks the truncation length + slice semantics.
+
+■ ast_import_guard (post-edit verification)
+- `python -c "import ast; ..."` on `src/phase_z2_verification_utils.py` → imports = `['__future__', 're', 'dataclasses', 'difflib', 'html.parser']`. Zero hits for `content_verifier` / `httpx` / `requests` / `openai` / `anthropic` / `kei`. Stdlib only; no AI/LLM/Kei/SSE/Phase Q reference path.
+
+■ test_run (u9 + u1~u8 regression guards)
+- `python -m pytest tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py -q` → `51 passed in 0.10s`. Per-unit count: u9 = 7 (this round) + u1-u8 = 44 (regression guards) = 51.
+
+■ scope_check
+- Changes for this unit are limited to `src/phase_z2_verification_utils.py` (append-only — u1~u8 surface untouched) and the new `tests/phase_z2/test_pz2_vu_invented.py`.
+- No edits to `orchestrator.py`, `src/content_verifier.py`, or any other Phase Z2 module.
+- No runtime wiring added (IMP-07 reverse-path gate respected — Step 1/2/14/21/22 wiring deferred to u11 design + IMP-16-U2 implementation).
+- `kw_ratio` Phase Q dead-branch simplification (drop `if keywords else 1.0`) is the only deviation from verbatim Phase Q text and is behavior-preserving.
+
+■ remaining_units (Stage 2 plan minus executed)
+- u10: sample-backed smoke test (no `orchestrator` import). Files: `tests/phase_z2/test_pz2_vu_integration.py`. Estimated 40 lines. Last runtime/test unit before u11.
+- u11: IMP-16-U2 design-only contract md (Step 1/2/14/21/22 wiring + redesigned frame-contract pattern dict + IMP-07 gate). Files: `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`. No tests (design-only). Estimated 50 lines.
+
+Executed so far (this Stage 3 cycle): u1, u2, u3, u4, u5, u6, u7, u8, u9.
+
+■ follow_up_issue_candidates
+- None this round. Scope-lock holds: H3 deterministic surface only; H4/H5 archive markers untouched; Phase Q `REQUIRED_PATTERNS` value reuse remains out of scope (slated for IMP-16-U2 redesign in u11). No new axes surfaced during u9 implementation.
+
+STOP. No commit, no push (Stage 3 rule).

.orchestrator/drafts/16_stage_code-edit_codex_r1.md

@@ -0,0 +1,44 @@
+[Codex #6] Stage 3 r1 verification for IMP-16-U1 unit u1.
+
+unit_verified
+- id: u1
+- covers: IMP-16-U1 module anchor and VerificationResult dataclass
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` is a new Phase Z-owned deterministic module anchor and exposes only the `VerificationResult` dataclass for u1.
+- The dataclass fields match the Phase Q source shape at `src/content_verifier.py` lines 29-36: `passed`, `area_name`, `checks`, `score`, `errors`, `warnings`.
+- Mutable fields use `field(default_factory=...)`; the u1 test verifies independent default collections.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; the substring appears only in docstrings/test assertions, and the AST import guard passes.
+- No AI/Kei/httpx/SSE path was introduced.
+
+scope_check
+- Changed files for this unit are limited to the claimed new module and u1 test file.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py'`
+- `python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py -q`
+- `rg "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_dataclass.py`
+- `(Get-Content -LiteralPath 'src\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_dataclass.py').Count`
+- `rg -n "class VerificationResult|passed:|area_name:|checks:|score:|errors:|warnings:" src/content_verifier.py src/phase_z2_verification_utils.py`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+Test results:
+- `python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py -q` -> `4 passed in 0.02s`
+Verified facts:
+- `src/phase_z2_verification_utils.py` is 30 lines.
+- `tests/phase_z2/test_pz2_vu_dataclass.py` is 59 lines.
+- AST import test confirms no `content_verifier` import.
+- `VerificationResult` required fields and default factories are covered by tests.
+- Only unit u1 is complete; u2 through u11 remain.
+Remaining units: [u2, u3, u4, u5, u6, u7, u8, u9, u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r10.md

@@ -0,0 +1,48 @@
+[Codex #24] Stage 3 r10 verification for IMP-16-U1 unit u10.
+
+unit_verified
+- id: u10
+- covers: sample-backed smoke without pipeline import (`tests/phase_z2/test_pz2_vu_integration.py`)
+
+verdict
+- Correct for the single executed unit.
+- `tests/phase_z2/test_pz2_vu_integration.py` adds a consumer-side integration smoke over a real `samples/mdx_batch/02.mdx` fixture without editing runtime code, which matches the Stage 2 u10 file scope.
+- The smoke exercises the deterministic chain through the public Phase Z surfaces: `verify_text_preservation(...) -> VerificationResult` and `detect_invented_text(...) -> list[str]`.
+- The mirrored HTML test verifies the expected preservation pass at the default threshold without asserting a sample-specific score or sentence count, so it does not make the pipeline depend on a hardcoded MDX 03/04/05 outcome.
+- The fabricated HTML test verifies the invented-text guard flags text with no source keyword anchor.
+- The AST import guard checks both `src/phase_z2_verification_utils.py` and the u10 test file for forbidden imports of `orchestrator`, `src.phase_z2_pipeline`, `src.content_verifier`, and `src.kei_client`; direct AST inspection confirmed only stdlib imports in the production module and the expected `src.phase_z2_verification_utils` import in the test.
+- The implementation remains deterministic and pure: no I/O beyond reading the local sample fixture in the test, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+
+scope_check
+- Changed files for this unit are limited to `tests/phase_z2/test_pz2_vu_integration.py`.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_integration.py'`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\samples\mdx_batch' | Select-Object Name,Length`
+- `rg -n "orchestrator|phase_z2_pipeline|content_verifier|kei_client|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_integration.py`
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_integration.py`
+- `python -m pytest tests/phase_z2/test_pz2_vu_integration.py -q`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_integration.py`
+- `rg -n "def (verify_text_preservation|detect_invented_text|extract_text_from_html|normalize_for_comparison|split_into_sentences)|class VerificationResult|def _sentence_matches_html" src/phase_z2_verification_utils.py`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_integration.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\samples\mdx_batch\02.mdx`
+Test results:
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_integration.py` passed
+- `python -m pytest tests/phase_z2/test_pz2_vu_integration.py -q` -> `5 passed in 0.11s`
+Verified facts:
+- u10 adds only the integration test file and does not modify runtime code.
+- The sample fixture `samples/mdx_batch/02.mdx` exists.
+- Production utility AST imports are `__future__`, `dataclasses`, `difflib`, `html.parser`, and `re`.
+- u10 test AST imports are `__future__`, `ast`, `pathlib`, and `src.phase_z2_verification_utils`.
+- Forbidden terms found by `rg` are only in docstrings/comments/constants that define the guard, not import statements.
+- The utility exposes the expected chain symbols: `VerificationResult`, `extract_text_from_html`, `normalize_for_comparison`, `split_into_sentences`, `_sentence_matches_html`, `verify_text_preservation`, and `detect_invented_text`.
+Remaining units: [u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r11.md

@@ -0,0 +1,51 @@
+[Codex #26] Stage 3 r11 verification for IMP-16-U2 unit u11.
+
+unit_verified
+- id: u11
+- covers: `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`, the design-only contract for Step 1 / 2 / 14 / 21 / 22 wiring, redesigned frame-contract reservation, trace/report hooks, and the IMP-07 hard gate.
+
+verdict
+- Correct for the single executed unit.
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` is documentation-only and adds no runtime code, no tests, and no pipeline wiring.
+- The document keeps U2 blocked behind IMP-07 by requiring IMP-07 to be `implemented` and `verified`, requiring a runtime reverse-path hit under `src/`, and requiring deterministic reverse-path outputs for Step 2 and Step 14.
+- The Step 1 / Step 2 / Step 14 / Step 21 / Step 22 contracts are additive and scoped to reverse-path re-entry. Original-upload behavior is explicitly unchanged.
+- Step 2 uses the U1 `verify_text_preservation` surface and the existing U1 default threshold; Step 14 uses `detect_invented_text` as telemetry only; Step 21 fields are additive; Step 22 is read-only/no-AI surface.
+- The redesigned frame-contract pattern section does not reuse Phase Q `REQUIRED_PATTERNS` values, reserves Phase Z frame-contract validation for IMP-20, and forbids importing or wrapping Phase Q `verify_structure`, `verify_area`, or `verify_all_areas`.
+- The guardrails preserve the Stage 1/2 scope lock: no AI/Kei/httpx/SSE path, no `src.content_verifier` import, no FORBIDDEN_KEI_MEMOS or `generate_with_retry` port, no sample-specific hardcoding, no spacing/CSS shrink, and no Step 20 status semantic change.
+- A runtime grep under `src/` found no `html_to_slide_mdx`, `edited_html_to_mdx`, `reverse_path`, or new re-entry debug-key wiring hits, so the IMP-07 gate remains respected.
+
+scope_check
+- Changed files for this unit are limited to `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction and because u11 is design-only with no tests.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-16-U2-WIRING-DESIGN.md'`
+- `rg -n "html_to_slide_mdx|edited_html_to_mdx|reverse_path|IMP-16|Step 1|Step 2|Step 14|Step 21|Step 22|src\.content_verifier|content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" docs/architecture/IMP-16-U2-WIRING-DESIGN.md src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_integration.py`
+- `Test-Path -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md'; Test-Path -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md'; Test-Path -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `$p='D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md'; $lines=Get-Content -LiteralPath $p; foreach ($i in 45..72) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `rg -n "Step 1|Step 2|Step 14|Step 21|Step 22|Operating Principles" docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md`
+- `rg -n "html_to_slide_mdx|edited_html_to_mdx|reverse_path|reentry_text_preservation|reentry_sentence_count|reentry_invented_text_fragments|reverse_path_reentry" src -g "*.py"`
+- `rg -n "REQUIRED_PATTERNS|verify_structure|verify_area|verify_all_areas|body_bg|core|sidebar|footer|frame_contracts\.yaml|IMP-20" docs/architecture/IMP-16-U2-WIRING-DESIGN.md src/phase_z2_verification_utils.py`
+- `Select-String -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md' -Pattern 'IMP-07|IMP-16|IMP-20' | ForEach-Object { '{0}: {1}' -f $_.LineNumber, $_.Line }`
+- `$p='D:\ad-hoc\kei\design_agent\docs\architecture\IMP-16-U2-WIRING-DESIGN.md'; (Get-Content -LiteralPath $p | Measure-Object -Line).Lines; Select-String -LiteralPath $p -Pattern 'line 51|:67|line 71|IMP-07 backlog|IMP-16 backlog|IMP-20' | ForEach-Object { '{0}: {1}' -f $_.LineNumber, $_.Line }`
+- `$p='D:\ad-hoc\kei\design_agent\docs\architecture\IMP-16-U2-WIRING-DESIGN.md'; $lines=Get-Content -LiteralPath $p; for ($i=1; $i -le $lines.Count; $i++) { if ($lines[$i-1] -match '^## |^### |^# ') { '{0}: {1}' -f $i, $lines[$i-1] } }`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\IMP-16-U2-WIRING-DESIGN.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_integration.py`
+Test results:
+- Not run. Current Stage 3 task requires verifying only u11; u11 is design-only and has no tests.
+Verified facts:
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` exists and contains the u11 design-only contract.
+- Backlog anchors exist for IMP-07 at line 51, IMP-16 at line 67, and IMP-20 at line 71.
+- `PHASE-Z-PIPELINE-OVERVIEW.md` contains anchors for Step 1, Step 2, Step 14, Step 21, Step 22, and Operating Principles.
+- `src/phase_z2_verification_utils.py` exists as the U1 surface consumed by the design.
+- Runtime grep under `src/` returned no hits for reverse-path implementation or re-entry debug-key wiring.
+- `content_verifier`, `FORBIDDEN_KEI_MEMOS`, `generate_with_retry`, `httpx`, `SSE`, `requests`, and `openai` references in the checked U11 scope are documentation/test guard references only, not new runtime wiring.
+Remaining units: none
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_code-edit_codex_r2.md

@@ -0,0 +1,50 @@
+[Codex #8] Stage 3 r2 verification for IMP-16-U1 unit u2.
+
+unit_verified
+- id: u2
+- covers: pure HTML text extraction (`_TextExtractor` + `extract_text_from_html`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds the stdlib-only `_TextExtractor(HTMLParser)` and `extract_text_from_html(html: str) -> list[str]` surface for u2.
+- The extractor behavior matches the Phase Q reference shape at `src/content_verifier.py` lines 43-66: ordered visible text fragments, stripped non-empty chunks, and skipped `<style>` / `<script>` bodies.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; references appear only in comments/docstrings/test assertions, and the u1 AST import guard remains green.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and the new u2 test file `tests/phase_z2/test_pz2_vu_extract.py`, with u1's existing dataclass test rerun as a guard.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_extract.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py'`
+- `python -m pytest tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_dataclass.py -q`
+- `rg -n "class _TextExtractor|def extract_text_from_html|handle_starttag|handle_endtag|handle_data" src/content_verifier.py src/phase_z2_verification_utils.py`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_dataclass.py`
+- `(Get-Content -LiteralPath 'src\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_extract.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_dataclass.py').Count`
+- `$lines = Get-Content -LiteralPath 'src\content_verifier.py'; for ($i=42; $i -le 70; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `$lines = Get-Content -LiteralPath 'src\phase_z2_verification_utils.py'; for ($i=34; $i -le 68; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_extract.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+Test results:
+- `python -m pytest tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_dataclass.py -q` -> `9 passed in 0.03s`
+Verified facts:
+- `src/phase_z2_verification_utils.py` is 71 lines.
+- `tests/phase_z2/test_pz2_vu_extract.py` is 54 lines.
+- `tests/phase_z2/test_pz2_vu_dataclass.py` is 59 lines.
+- `_TextExtractor` skips `style` and `script`, strips non-empty data chunks, drops whitespace-only chunks, and preserves fragment order.
+- `extract_text_from_html` is a pure wrapper around `_TextExtractor().feed(html)` and returns `parser.texts`.
+- u2 tests cover order preservation, style/script exclusion, whitespace stripping, Korean/inline markup fragments, and empty input.
+- The u1 AST import guard confirms no `content_verifier` import remains.
+- Only units u1 and u2 are complete; u3 through u11 remain.
+Remaining units: [u3, u4, u5, u6, u7, u8, u9, u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r3.md

@@ -0,0 +1,52 @@
+[Codex #10] Stage 3 r3 verification for IMP-16-U1 unit u3.
+
+unit_verified
+- id: u3
+- covers: Korean text normalization constants and function (`_PARTICLES`, `_ENDING_NORMALIZE`, `normalize_for_comparison`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds stdlib-only `re`, `_PARTICLES`, `_ENDING_NORMALIZE`, and `normalize_for_comparison(text: str) -> str` for u3.
+- The Phase Z normalization pipeline matches the Phase Q reference shape at `src/content_verifier.py` lines 78-106: collapse whitespace, strip bullet markers with `r"[•◦·\-▪▸►]"`, decode the same small HTML entity set, then fold one trailing ending using `_ENDING_NORMALIZE`.
+- `_PARTICLES` matches the Phase Q Korean-locale particle inventory and is sorted longest-first for downstream greedy keyword stripping.
+- `_ENDING_NORMALIZE` has the same effective 7-key mapping as the Phase Q source after Python duplicate-key resolution.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module, with `content_verifier` appearing only in docstrings/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and the new u3 test file `tests/phase_z2/test_pz2_vu_normalize.py`, with u1/u2 tests rerun as guards.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_normalize.py'`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=78; $i -le 120; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m pytest tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py`
+- `rg -n '_PARTICLES|_ENDING_NORMALIZE|def normalize_for_comparison|re\.sub|text\.replace|text\.endswith' src/content_verifier.py src/phase_z2_verification_utils.py`
+- `(Get-Content -LiteralPath 'src\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_normalize.py').Count`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_normalize.py`
+- AST literal comparison of `_PARTICLES` and `_ENDING_NORMALIZE` in `src/content_verifier.py` and `src/phase_z2_verification_utils.py`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_normalize.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_extract.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+Test results:
+- `python -m pytest tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py -q` -> `15 passed in 0.04s`
+Verified facts:
+- `src/phase_z2_verification_utils.py` is 103 lines.
+- `tests/phase_z2/test_pz2_vu_normalize.py` is 64 lines.
+- `_PARTICLES` has 20 entries, matches the Phase Q literal inventory, and is sorted longest-first.
+- `_ENDING_NORMALIZE` has the same effective 7-entry mapping as Phase Q: `있음`, `됨`, `함`, `임`, `없음`, `았음`, `었음`.
+- `normalize_for_comparison` collapses whitespace, strips the Phase Q bullet marker class, decodes `&`, `<`, `>`, ` `, `'`, and `"`, and folds only one trailing ending.
+- AST import scan confirms no `content_verifier` import in the Phase Z utility module.
+- Only unit u3 is newly verified in this turn; u4 through u11 remain.
+Remaining units: [u4, u5, u6, u7, u8, u9, u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r4.md

@@ -0,0 +1,49 @@
+[Codex #12] Stage 3 r4 verification for IMP-16-U1 unit u4.
+
+unit_verified
+- id: u4
+- covers: keyword extraction (`extract_keywords`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds `extract_keywords(text: str) -> list[str]` for u4.
+- The keyword extraction behavior matches the Phase Q reference surface at `src/content_verifier.py` lines 120-133: `re.findall(r"[가-힣a-zA-Z0-9()]+", text)`, drop tokens shorter than 3 characters, strip one longest-match trailing particle when the remaining stem is length >= 2, then append tokens whose final length is >= 2.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module plus the expected test import from `src.phase_z2_verification_utils`. `content_verifier` appears only in docstrings/comments/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and the new u4 test file `tests/phase_z2/test_pz2_vu_keywords.py`, with u1-u3 tests rerun as guards.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_keywords.py'`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=78; $i -le 145; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m pytest tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py`
+- `rg -n "def extract_keywords|re\.findall|_PARTICLES|len\(w\)|keywords\.append" src/content_verifier.py src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_keywords.py`
+- `(Get-Content -LiteralPath 'src\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_keywords.py').Count`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_keywords.py`
+- AST dump comparison for `extract_keywords` in `src/content_verifier.py` and `src/phase_z2_verification_utils.py`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_keywords.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_extract.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_normalize.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+Test results:
+- `python -m pytest tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py -q` -> `21 passed in 0.05s`
+Verified facts:
+- `src/phase_z2_verification_utils.py` is 125 lines.
+- `tests/phase_z2/test_pz2_vu_keywords.py` is 52 lines.
+- AST import scan found no `content_verifier` import in the Phase Z utility module or u4 test file.
+- `extract_keywords` in the Phase Z utility uses the same operative AST shape as the Phase Q reference, with only type annotation/docstring differences.
+- Unit u4 is complete; u5 through u11 remain.
+Remaining units: [u5, u6, u7, u8, u9, u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r5.md

@@ -0,0 +1,48 @@
+[Codex #14] Stage 3 r5 verification for IMP-16-U1 unit u5.
+
+unit_verified
+- id: u5
+- covers: meta-line stripping (`_META_PREFIXES`, `_META_INLINE_FRAGMENTS`, `strip_meta_lines`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds `_META_PREFIXES`, `_META_INLINE_FRAGMENTS`, and `strip_meta_lines(text: str) -> str` for u5.
+- The meta-line stripping behavior matches the Phase Q reference surface at `src/content_verifier.py` lines 126-159: split on newline, strip only for matching, drop lines whose stripped form starts with a meta prefix, drop lines containing expression-hint fragments, preserve original whitespace for retained lines, and join with newline.
+- Lifting the three inline expression-hint fragments into `_META_INLINE_FRAGMENTS` keeps the same deterministic filter surface while making the Phase Z constants auditable and unit-testable.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module plus expected test imports from `src.phase_z2_verification_utils`. `content_verifier` appears only in docstrings/comments/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and the new u5 test file `tests/phase_z2/test_pz2_vu_meta_strip.py`, with u1-u4 tests rerun as guards.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_meta_strip.py'`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=120; $i -le 160; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=136; $i -le 172; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `python -m pytest tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `(Get-Content -LiteralPath 'src\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_meta_strip.py').Count`
+- `rg -n "_META_PREFIXES|_META_INLINE_FRAGMENTS|def strip_meta_lines|startswith|fragment in stripped|filtered\.append|join\(filtered\)" src/content_verifier.py src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `git diff -- src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_meta_strip.py`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_meta_strip.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_dataclass.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_extract.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_normalize.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_keywords.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+Test results:
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_meta_strip.py` passed.
+- `python -m pytest tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py -q` passed: 26 passed.
+Remaining units: [u6, u7, u8, u9, u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r6.md

@@ -0,0 +1,44 @@
+[Codex #16] Stage 3 r6 verification for IMP-16-U1 unit u6.
+
+unit_verified
+- id: u6
+- covers: sentence splitting (`split_into_sentences`, `_BULLET_MARKER_PATTERN`, `_SENTENCE_SPLIT_PATTERN`, `_MIN_SENTENCE_LEN`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds `split_into_sentences(text: str) -> list[str]` plus named module constants for the bullet marker regex, sentence split regex, and minimum sentence length.
+- The sentence splitting behavior matches the Phase Q reference surface at `src/content_verifier.py` lines 162-187: apply `strip_meta_lines` first, split by newline, strip each line, skip empty and `#`-led header lines, remove leading bullet/numeric markers with `^[\-•◦·\d]+[.)]\s*`, split on `(?<=\.)\s+`, and keep only parts with length at least 5.
+- Lifting the two regex literals and the length threshold into named constants preserves the deterministic behavior while making the Phase Z surface auditable and unit-testable.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module plus the expected test import from `src.phase_z2_verification_utils`. `content_verifier` appears only in docstrings/comments/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and the new u6 test file `tests/phase_z2/test_pz2_vu_sentence_split.py`, with u1-u5 tests rerun as guards.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_sentence_split.py'`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=170; $i -le 205; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=150; $i -le 188; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_sentence_split.py`
+- `python -m pytest tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_sentence_split.py`
+- `rg -n "def split_into_sentences|_BULLET_MARKER_PATTERN|_SENTENCE_SPLIT_PATTERN|_MIN_SENTENCE_LEN|strip_meta_lines\(|re\.sub|re\.split|len\(part\)" src/content_verifier.py src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_sentence_split.py`
+- `(Get-Content -LiteralPath 'src\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'tests\phase_z2\test_pz2_vu_sentence_split.py').Count`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_sentence_split.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+Test results:
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_sentence_split.py` passed.
+- `python -m pytest tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py -q` passed: 32 passed in 0.06s.
+Remaining units: [u7, u8, u9, u10, u11]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r7.md

@@ -0,0 +1,50 @@
+[Codex #18] Stage 3 r7 verification for IMP-16-U1 unit u7.
+
+unit_verified
+- id: u7
+- covers: private sentence-vs-html match helper (`_sentence_matches_html`, `_SENTENCE_KEYWORD_MATCH_THRESHOLD`, `_SENTENCE_SEQUENCE_MATCH_THRESHOLD`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds stdlib `SequenceMatcher`, the two named thresholds, and `_sentence_matches_html(sentence, html_combined, html_texts) -> bool`.
+- The helper matches the Phase Q reference decision body inside `verify_text_preservation`: normalize the original sentence, extract keywords, treat empty keyword sets as preserved, compute keyword ratio against pre-normalized combined HTML text, compute the best SequenceMatcher ratio across normalized HTML fragments, and accept on `kw_ratio >= 0.6` or `best_ratio >= 0.65`.
+- Lifting the `0.6` and `0.65` literals into named constants preserves behavior while making the Phase Z verification surface auditable.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module and the expected test import from `src.phase_z2_verification_utils`. `content_verifier` appears only in docstrings/comments/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_match_helper.py`.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\\ad-hoc\\kei\\design_agent\\src\\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\\ad-hoc\\kei\\design_agent\\tests\\phase_z2\\test_pz2_vu_match_helper.py'`
+- `$lines = Get-Content -LiteralPath 'D:\\ad-hoc\\kei\\design_agent\\src\\content_verifier.py'; for ($i=220; $i -le 260; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `$lines = Get-Content -LiteralPath 'D:\\ad-hoc\\kei\\design_agent\\src\\content_verifier.py'; for ($i=206; $i -le 236; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_match_helper.py`
+- `python -m pytest tests/phase_z2/test_pz2_vu_match_helper.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_match_helper.py`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_match_helper.py`
+- `rg -n "SequenceMatcher|_SENTENCE_KEYWORD_MATCH_THRESHOLD|_SENTENCE_SEQUENCE_MATCH_THRESHOLD|def _sentence_matches_html|kw_found|kw_ratio|best_ratio" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_match_helper.py src/content_verifier.py`
+- `(Get-Content -LiteralPath 'D:\\ad-hoc\\kei\\design_agent\\src\\phase_z2_verification_utils.py').Count; (Get-Content -LiteralPath 'D:\\ad-hoc\\kei\\design_agent\\tests\\phase_z2\\test_pz2_vu_match_helper.py').Count`
+- `rg -n "phase_z2_verification_utils|_sentence_matches_html|verify_text_preservation|detect_invented_text" src tests orchestrator.py docs -g "*.py" -g "*.md"`
+- `git diff -- src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_match_helper.py`
+Files checked:
+- `D:\\ad-hoc\\kei\\design_agent\\src\\phase_z2_verification_utils.py`
+- `D:\\ad-hoc\\kei\\design_agent\\tests\\phase_z2\\test_pz2_vu_match_helper.py`
+- `D:\\ad-hoc\\kei\\design_agent\\src\\content_verifier.py`
+Test results:
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_match_helper.py` passed.
+- `python -m pytest tests/phase_z2/test_pz2_vu_match_helper.py -q` passed: 5 passed in 0.02s.
+Verified facts:
+- `_sentence_matches_html` computes `norm_orig`, `keywords`, `kw_found`, `kw_ratio`, per-fragment `SequenceMatcher` ratios, and returns true on the locked keyword or sequence thresholds.
+- `_SENTENCE_KEYWORD_MATCH_THRESHOLD == 0.6` and `_SENTENCE_SEQUENCE_MATCH_THRESHOLD == 0.65` are locked by unit tests.
+- Scope search found no runtime import or pipeline wiring for `phase_z2_verification_utils`; usage is limited to the Phase Z2 tests and the new helper itself.
+- Forbidden-token search found only explanatory docstring/comment occurrences, not runtime imports or calls.
+Remaining units: [u8, u9, u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r8.md

@@ -0,0 +1,53 @@
+[Codex #20] Stage 3 r8 verification for IMP-16-U1 unit u8.
+
+unit_verified
+- id: u8
+- covers: `verify_text_preservation` (`_TEXT_PRESERVATION_DEFAULT_THRESHOLD`, `_MISSING_SENTENCE_REPORT_LIMIT`, `_MISSING_SENTENCE_TRUNCATE_LEN`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds the Phase Z-owned deterministic `verify_text_preservation(original_mdx, generated_html, area_name, threshold=...) -> VerificationResult` surface plus named constants for the Phase Q default threshold and missing-sentence report limits.
+- The implementation matches the Phase Q reference at `src/content_verifier.py:206-257`: split original MDX into sentences, return passed/score 1.0 when no sentences exist, extract visible HTML text, normalize the joined HTML once, count sentence matches, compute `matched / len(original_sentences)`, compare against the threshold, report capped missing sentences, and attach a preservation warning when score is below 1.0.
+- The per-sentence matching logic is correctly delegated to the already-verified u7 helper `_sentence_matches_html`, preserving the Phase Q keyword-ratio and SequenceMatcher semantics without duplicating that body in u8.
+- Lifting `0.70`, `5`, and `60` into named constants preserves behavior while making the Phase Z verification surface auditable and unit-testable.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module and the expected test import from `src.phase_z2_verification_utils`. `content_verifier` appears only in docstrings/comments/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_preservation.py`.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_preservation.py'`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=190; $i -le 257; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_preservation.py`
+- `python -m pytest tests/phase_z2/test_pz2_vu_preservation.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_preservation.py`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_preservation.py`
+- `rg -n "def verify_text_preservation|threshold: float = 0.70|missing\[:5\]|s\[:60\]|보존율|누락 문장" src/content_verifier.py src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_preservation.py`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_preservation.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+
+Test results:
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_preservation.py` passed.
+- `python -m pytest tests/phase_z2/test_pz2_vu_preservation.py -q` passed: 7 passed in 0.02s.
+
+Verified facts:
+- Phase Q reference default threshold is `0.70`; Phase Z u8 exposes `_TEXT_PRESERVATION_DEFAULT_THRESHOLD = 0.70` and uses it as the default.
+- Phase Q reference caps missing sentence display at `missing[:5]` and truncates with `s[:60]`; Phase Z u8 exposes `_MISSING_SENTENCE_REPORT_LIMIT = 5` and `_MISSING_SENTENCE_TRUNCATE_LEN = 60` and uses both in the error renderer.
+- Phase Z u8 returns `VerificationResult(passed=True, checks={"text_preservation": True}, score=1.0)` when the original sentence list is empty, matching the Phase Q early-return behavior.
+- Phase Z u8 computes warnings only when score is below 1.0, matching the Phase Q warning surface.
+- Import scan found no `src.content_verifier` import and no non-stdlib runtime imports in `src/phase_z2_verification_utils.py`.
+- Forbidden terms scan found `content_verifier`, `LLM`, `httpx`, and `SSE` only in module/test documentation text, not as imports or executable calls.
+Remaining units: [u9, u10, u11]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_code-edit_codex_r9.md

@@ -0,0 +1,54 @@
+[Codex #22] Stage 3 r9 verification for IMP-16-U1 unit u9.
+
+unit_verified
+- id: u9
+- covers: `detect_invented_text` (`_INVENTED_TEXT_MIN_LENGTH`, `_INVENTED_TEXT_ALLOWED_LABELS`, `_INVENTED_TEXT_CSS_NUMBER_PATTERN`, `_INVENTED_TEXT_KEYWORD_THRESHOLD`, `_INVENTED_TEXT_TRUNCATE_LEN`)
+
+verdict
+- Correct for the single executed unit.
+- `src/phase_z2_verification_utils.py` adds the Phase Z-owned deterministic `detect_invented_text(original_mdx, generated_html, min_length=...) -> list[str]` surface plus named constants for the Phase Q minimum length, structural label allow-list, CSS/numeric skip regex, keyword-overlap threshold, and report truncation length.
+- The implementation matches the Phase Q reference at `src/content_verifier.py:260-298`: extract visible HTML text, normalize the original MDX once, skip short fragments, skip structural labels, skip CSS/numeric fragments, normalize each surviving fragment, extract keywords, skip empty keyword sets, compute keyword coverage against normalized MDX, and flag fragments whose keyword ratio is below `0.4`, truncating flagged values to 80 characters.
+- Lifting `15`, the three allowed labels, `r"^[\d\s.,%px#rgb()]+$"`, `0.4`, and `80` into named constants preserves behavior while making the Phase Z verification surface auditable and unit-testable.
+- The implementation remains deterministic and pure: no I/O, no network call, no AI/Kei path, no `httpx`/SSE path.
+- No runtime wiring was added, so the IMP-07 gate remains respected.
+- No direct `src.content_verifier` import is present; AST import checks show only stdlib imports in the module and the expected test import from `src.phase_z2_verification_utils`. `content_verifier` appears only in docstrings/comments/test text.
+
+scope_check
+- Changed files for this unit are limited to `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_invented.py`.
+- Existing unrelated worktree changes are present and were not touched.
+- I did not run full pytest, per the current Stage 3 instruction to verify only the single executed unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py'`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_invented.py'`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=276; $i -le 315; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `$lines = Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\content_verifier.py'; for ($i=260; $i -le 315; $i++) { if ($i -le $lines.Count) { '{0}: {1}' -f $i, $lines[$i-1] } }`
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_invented.py`
+- `python -m pytest tests/phase_z2/test_pz2_vu_invented.py -q`
+- `rg -n "content_verifier|httpx|generate_with_retry|FORBIDDEN_KEI_MEMOS|LLM|SSE|requests|openai" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_invented.py`
+- AST import scan for `src/phase_z2_verification_utils.py` and `tests/phase_z2/test_pz2_vu_invented.py`
+- `rg -n "_INVENTED_TEXT_|def detect_invented_text|kw_found|kw_ratio|invented\.append|extract_text_from_html\(|normalize_for_comparison\(|extract_keywords\(" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_invented.py`
+- `git diff -- src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_invented.py`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\tests\phase_z2\test_pz2_vu_invented.py`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+
+Test results:
+- `python -m py_compile src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_invented.py` passed.
+- `python -m pytest tests/phase_z2/test_pz2_vu_invented.py -q` passed: 7 passed in 0.03s.
+
+Verified facts:
+- `detect_invented_text` composes existing Phase Z utilities: `extract_text_from_html`, `normalize_for_comparison`, and `extract_keywords`.
+- The function returns `list[str]` and has default `min_length = _INVENTED_TEXT_MIN_LENGTH`.
+- The constants are locked by `tests/phase_z2/test_pz2_vu_invented.py`.
+- The focused u9 tests cover preserved text, low keyword overlap, short-fragment skip, allowed-label skip, CSS/numeric skip, and 80-character truncation.
+- `rg` hits for guarded terms are limited to docstrings/comments/test reference text; no forbidden runtime import or AI/network path was found.
+- AST import scan found only `__future__`, `re`, `dataclasses`, `difflib`, and `html.parser` in the module, plus the expected test import from `src.phase_z2_verification_utils`.
+- No Step 1/2/14/21/22 pipeline wiring was added.
+Remaining units: [u10, u11]
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_commit-push_claude_r1.md

@@ -0,0 +1,48 @@
+📌 **[오케스트레이터] Stage 5: 커밋 및 푸쉬 완료**
+
+■ **핵심 결정 (Korean)**
+- Commit `23ba8b6` (`feat(IMP-16): U1 H3 verification utility port + U2 wiring design`) 생성. 정확히 12 파일만 staged — `src/phase_z2_verification_utils.py` (+335 new), `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` (+75 new), `tests/phase_z2/test_pz2_vu_*.py` × 10 (+746 new). 총 1156 insertions.
+- `git add` 는 12 파일을 명시적으로만 호출 — `orchestrator.py` / `tests/orchestrator_unit/test_subprocess_cleanup.py` (별개 dirty M) + 모든 다른 untracked (.orchestrator/, chromedriver/, scripts/, templates/catalog/, tests/matching/, samples/uploads/, 등) 는 staged 에서 제외 확인 (`git diff --staged --stat` = 12 파일, 1156 insertions, 0 deletions).
+- Pre-commit hook 통과 (no `--no-verify`, no `-c commit.gpgsign=false`). Commit body 가 U1 axis (H3 utility port, deterministic, stdlib only) 와 u11 (design-only md) 를 binding 으로 명시, IMP-07 hard gate / Phase Q forbidden-coupling 가드레일 재확인.
+- Push : `origin` (GitHub `keimin86/design_agent`) + `slide2` (Gitea `Kyeongmin/C.E.L_Slide_test2`, issue #16 의 canonical 원격) 둘 다 `614c533..23ba8b6` 로 fast-forward.
+- Remote 검증 — `git ls-remote slide2 refs/heads/main` 과 `git ls-remote origin refs/heads/main` 모두 `23ba8b68cdc315cbab542ce8c285caf2f431dbe5` 반환, local `HEAD` 와 일치.
+
+■ **범위 제외 (커밋 scope 와 일치)**
+- `orchestrator.py` / `tests/orchestrator_unit/test_subprocess_cleanup.py` 수정분 (prior session) — staged X, working tree 에 그대로 유지.
+- `.orchestrator/`, `chromedriver/`, `scripts/post_gitea_comment.py`, `scripts/sync_phase_z_run_to_frontend.py`, `templates/catalog/`, `templates/phase_z2/families/*.html`, `templates/phase_z2/frames/`, `tests/matching/*`, `tests/pipeline/`, `tests/CLAUDE.md`, `tests/PIPELINE.md`, `tests/PLAN.md`, `tests/PROGRESS.md`, `tests/README.md` untracked — staged X.
+- `samples/uploads/`, `samples/mdx_batch/05.mdx`, `samples/mdx/05. 설계 방식의 왜곡.mdx`, `orchestrator(원본).py` untracked — staged X.
+- U2 runtime wiring (Step 1 / 2 / 14 / 21 / 22 reverse-path re-entry) — 코드 변경 없음. u11 는 design-only md 로 IMP-07 hard gate 보존.
+
+■ **다음 단계** = Stage 6 (final-close) — issue #16 의 Codex independent verification + close 권한 위임 또는 별도 final-close turn.
+
+=== EVIDENCE ===
+
+**commands_run**:
+- `git status --short` → confirmed 12 IMP-16 in-scope untracked items + 2 out-of-scope M (orchestrator.py, tests/orchestrator_unit/test_subprocess_cleanup.py) + many out-of-scope untracked items.
+- `git add src/phase_z2_verification_utils.py docs/architecture/IMP-16-U2-WIRING-DESIGN.md tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_integration.py` (explicit 12-file list, no `-A` / `.`).
+- `git diff --staged --stat` → 12 files, 1156 insertions, 0 deletions.
+- `git status --short` (post-add) → exactly 12 `A` entries + the 2 out-of-scope `M` entries left unchanged + all other items still `??`.
+- `git commit -m <<HEREDOC>>` → `[main 23ba8b6]` produced; 12 files changed, 1156 insertions(+).
+- `git remote -v` → `origin` = GitHub `keimin86/design_agent`, `slide2` = Gitea `Kyeongmin/C.E.L_Slide_test2` (canonical for issue #16, per `[reference_design_agent_remotes]`), `gitea` = unrelated `_Geulbeot`.
+- `git push origin main` → `614c533..23ba8b6  main -> main`.
+- `git push slide2 main` → `614c533..23ba8b6  main -> main`, Gitea processed 1 reference.
+- `git ls-remote origin refs/heads/main` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5` (matches HEAD).
+- `git ls-remote slide2 refs/heads/main` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5` (matches HEAD).
+- `git rev-parse HEAD` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+
+**facts_checked**:
+- commit scope = exactly the 12 files listed in Stage 4 binding exit report (U1 runtime u1~u10 + u11 design md). No bundled unrelated dirty state.
+- pre-commit hooks ran clean (no `--no-verify`).
+- both push targets converged on the same SHA as local HEAD.
+- prior session's dirty files (`orchestrator.py` M, `tests/orchestrator_unit/test_subprocess_cleanup.py` M) still M in working tree — untouched by this commit, as required by Stage 4 binding scope.
+- no `src/content_verifier` reference, no `FORBIDDEN_KEI_MEMOS` / `generate_with_retry` / `REQUIRED_PATTERNS` / `verify_structure` / `verify_area` / `verify_all_areas` runtime usage in the staged diff (verified at Stage 4, unchanged at commit).
+- IMP-07 hard gate preserved — no runtime wiring of Step 1 / 2 / 14 / 21 / 22 reverse-path lands in this commit; u11 ships as `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` design-only contract.
+
+**commit_sha**: `23ba8b68cdc315cbab542ce8c285caf2f431dbe5` (short `23ba8b6`).
+**parent**: `614c533 feat(IMP-15): 실행-4 — debug.json event surfacing + spec taxonomy row` (baseline at Stage 1 entry).
+**remotes_synced**: origin (GitHub `keimin86/design_agent`) + slide2 (Gitea `Kyeongmin/C.E.L_Slide_test2`).
+
+**source_comment_ids**: BASIS [Claude #5] (Stage 4 PASS exit report), [Codex #27] (Stage 4 independent YES).
+
+rewind_target: n/a
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_commit-push_codex_r1.md

@@ -0,0 +1,61 @@
+[Codex #28] Stage 5 independent commit/push verification for IMP-16.
+
+Decision: PASS. Commit `23ba8b68cdc315cbab542ce8c285caf2f431dbe5` is local `HEAD` on branch `main`, and both pushed remotes checked for this stage (`origin/main` and `slide2/main`) resolve to the same SHA.
+
+Commit scope verified: the commit adds exactly the 12 IMP-16 files claimed by the Stage 5 report:
+- `src/phase_z2_verification_utils.py`
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`
+- `tests/phase_z2/test_pz2_vu_dataclass.py`
+- `tests/phase_z2/test_pz2_vu_extract.py`
+- `tests/phase_z2/test_pz2_vu_normalize.py`
+- `tests/phase_z2/test_pz2_vu_keywords.py`
+- `tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `tests/phase_z2/test_pz2_vu_sentence_split.py`
+- `tests/phase_z2/test_pz2_vu_match_helper.py`
+- `tests/phase_z2/test_pz2_vu_preservation.py`
+- `tests/phase_z2/test_pz2_vu_invented.py`
+- `tests/phase_z2/test_pz2_vu_integration.py`
+
+No unintended tracked or untracked working-tree files were included in the commit. The dirty working tree remains outside commit scope: `orchestrator.py` and `tests/orchestrator_unit/test_subprocess_cleanup.py` are still modified but unstaged; the large untracked local artifact set remains untracked. `git diff --cached --stat` is empty after the commit.
+
+No full pytest was run in this Stage 5 verification, per instruction. Stage 5 verification was limited to commit SHA, committed file scope, staging state, and remote reflection.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `git rev-parse HEAD`
+- `git log -1 --pretty=fuller --stat`
+- `git remote -v`
+- `git branch --show-current`
+- `git show --name-status --format=fuller --no-renames HEAD`
+- `git diff --cached --stat`
+- `git ls-remote origin refs/heads/main`
+- `git ls-remote slide2 refs/heads/main`
+
+Files checked:
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`
+- `src/phase_z2_verification_utils.py`
+- `tests/phase_z2/test_pz2_vu_dataclass.py`
+- `tests/phase_z2/test_pz2_vu_extract.py`
+- `tests/phase_z2/test_pz2_vu_integration.py`
+- `tests/phase_z2/test_pz2_vu_invented.py`
+- `tests/phase_z2/test_pz2_vu_keywords.py`
+- `tests/phase_z2/test_pz2_vu_match_helper.py`
+- `tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `tests/phase_z2/test_pz2_vu_normalize.py`
+- `tests/phase_z2/test_pz2_vu_preservation.py`
+- `tests/phase_z2/test_pz2_vu_sentence_split.py`
+
+Commit SHA: `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`
+
+Verified facts:
+- Local branch is `main`.
+- Local `HEAD` is `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `origin/main` resolves to `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `slide2/main` resolves to `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- Commit stat is 12 files changed, 1156 insertions, 0 deletions.
+- `git show --name-status --format=fuller --no-renames HEAD` lists only 12 added files, all in the IMP-16 scoped path set above.
+- `git diff --cached --stat` is empty, so no files remain staged after commit.
+- Out-of-scope local dirty files remain outside the commit.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_final-close_claude_r1.md

@@ -0,0 +1,108 @@
+[Claude #6] Stage 6 — final-close Round #1 for IMP-16.
+
+## 1. Issue body re-read (binding scope)
+
+- Issue #16 = IMP-16 B-2 verification helper axis (Step 1/2/14/21/22 helper layer).
+- `hard link: IMP-07` — IMP-07 was the parent reverse path (B-2 main).
+- Stage 1 split: U1 = port (free-to-land), U2 = wiring (gated by IMP-07).
+- Review-loop boxes mapped to stages: Codex #1 (Stage 1) → Claude #4 + Codex #3 (Stage 2) → Claude #5 + Codex #27 (Stage 4) → Claude (Stage 3 commit) + Codex #28 (Stage 5).
+
+## 2. Commit verification on origin
+
+- Local HEAD = `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `git ls-remote origin refs/heads/main` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `git ls-remote slide2 refs/heads/main` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5` (design_agent Gitea mirror per `reference_design_agent_remotes.md`).
+- `git show --stat 23ba8b6` → 12 files, all additions, 1156 insertions, 0 deletions.
+- Commit message: `feat(IMP-16): U1 H3 verification utility port + U2 wiring design`.
+- Untracked working-tree (orchestrator.py / matching artifacts / etc.) untouched — Stage 3 contract preserved.
+
+## 3. Goal vs result
+
+| Axis | Issue goal | Result | Status |
+|---|---|---|---|
+| H3 utility port to Phase Z-owned module | Required (Stage 1 U1) | `src/phase_z2_verification_utils.py` (335 LOC, stdlib only) | ✅ delivered |
+| Deterministic surface, no AI/Kei | Required guardrail | re / dataclasses / difflib / html.parser only; no httpx / no SSE / no LLM | ✅ delivered |
+| No Phase Q coupling | Required guardrail | No `src.content_verifier` import; no `REQUIRED_PATTERNS` / `verify_structure` / `verify_area` reuse; no `FORBIDDEN_KEI_MEMOS` / `generate_with_retry` port | ✅ delivered |
+| Unit test 1:1 mapping | Stage 2 plan u1~u10 | 10 test files / 56 tests / 0.19s green | ✅ delivered |
+| Regression baseline | Stage 2 plan | 15 tests / 7.59s green | ✅ delivered |
+| U2 wiring (Step 1/2/14/21/22) | **Gated by IMP-07** (Stage 1 scope-lock) | Design-only md `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`, no runtime wiring | ✅ matches binding contract |
+
+Round #1 deliverable as defined by Stage 1 binding contract = met.
+
+## 4. Stage 1 assumption needing reconciliation (IMP-07 state)
+
+- Stage 1 binding contract (line: `IMP-07 = pending`) and Stage 4 evidence presumed IMP-07 was still pending.
+- Actual state at the time of Stage 1 (and at Stage 6 audit): IMP-07 was **closed at 2026-05-15T18:28** (4 days before IMP-16 commit) — closed as **not-needed**, not as implemented.
+  - User closing rationale (IMP-07 결과보고서 v2): "MDX → HTML 변환 이후 HTML 수기 수정은 HTML에서만 적용" — reverse path is not part of the current product direction.
+  - IMP-07 was closed without any reverse-path code landing — `html_to_slide_mdx` / `edited_html_to_mdx` / `reverse_path` / `reentry_*` are still absent from `src/` (re-verified at Stage 6).
+- Implication for IMP-16-U2: the hard-link parent is **retired**, not pending. U2 runtime wiring (Step 1/2/14/21/22 re-entry) has no production trigger to wire into. The design-only md becomes a historical archive of the would-have-been contract.
+- Implication for IMP-16-U1: still useful on its own as a Phase Z-owned deterministic text verification utility (text preservation + invented-text detection are useful outside the reverse path too — they can serve future Step 2 forward-path text guards). Code lands as planned; no rollback warranted.
+
+This is a **scope-context finding**, not a Round #1 regression — Round #1 delivered exactly what the binding contract specified. But the binding contract was written against a stale parent-status assumption.
+
+## 5. Follow-up links
+
+- IMP-07 (#7) closed as not-needed on 2026-05-15 → permanently gates IMP-16-U2. No reactivation expected unless reverse-path policy reverses.
+- IMP-15 — Step 14 image/table overflow (separate axis, in progress).
+- IMP-20 — frame-contract runtime pattern dict (deferred, parked in u11 md §future).
+- u3 `_PARTICLES` / `_ENDING_NORMALIZE` micro-axis (Stage 4 r1 micro-finding) — not in Round #1 scope.
+
+## 6. Labels
+
+- Current label: `needs-codex-review`.
+- After Stage 6 close: this label should be removed; Round #1 is fully reviewed (Stage 4 Codex #27 + Stage 5 Codex #28 both PASS).
+
+## 7. Close decision
+
+**Round #1 close: YES.**
+
+- Commit landed on `origin/main` + `slide2/main` with exact 12-file scope.
+- All Stage 1 guardrails preserved.
+- All u1~u11 implementation_units delivered.
+- All Stage 2 baseline + targeted tests green.
+- Stages 1→5 each ended with PASS and binding exit reports.
+
+**Issue #16 final disposition: defer to user.** Two consistent options (both supported by the new IMP-07 fact, both don't touch Round #1):
+
+- (a) Close IMP-16 entirely. U1 = landed; U2 = archive (parent retired). Mark U2 design md as historical reference.
+- (b) Keep IMP-16 open as "U2 archived, awaiting reverse-path policy reversal." Same code state; just different ticket-state.
+
+Per the 1-turn-1-step rule, I'm not picking between (a) and (b) here; surfacing them for user decision is Stage 6's scope.
+
+## 8. Unresolved questions
+
+- Final issue-state disposition (close vs keep open) — user decision.
+- Label cleanup policy (`needs-codex-review` → done / archive / something else) — user decision.
+- u3 KR normalization edge-axis follow-up — separate IMP, not Round #1.
+
+## 9. Guardrails (Stage 6 axis)
+
+- No source edits, no test reruns required at Stage 6 (commit/push verification axis only).
+- No retroactive scope expansion; Stage 1 binding contract honored.
+- IMP-07 status finding reported as fact, no auto-action taken on issue state / labels.
+
+=== EVIDENCE ===
+
+Commands run:
+
+- `git rev-parse HEAD` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `git log -1 --pretty=fuller --stat 23ba8b6` → 12 files / 1156 insertions / 0 deletions; commit message verbatim verified.
+- `git show --stat --name-status 23ba8b6` → all 12 paths marked `A` (added), exactly matches Stage 5 staged_files list.
+- `git ls-remote origin refs/heads/main` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `git ls-remote slide2 refs/heads/main` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`.
+- `git status --short` → no IMP-16 paths in unstaged/untracked output (12-file scope fully committed); pre-existing dirty entries (`orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py`, and many unrelated `tests/matching/*` untracked artifacts) remain unchanged — Stage 3 contract preserved.
+- `curl -s -H "Authorization: token $GITEA_TOKEN" .../issues/7` → IMP-07 `state=closed`, `closed_at=2026-05-15T18:28:24+09:00`.
+- `curl -s .../issues/7/comments` (last 2 comments) → "결과보고서 v2" confirms IMP-07 was closed as not-needed without any code landing.
+- `curl -s .../issues/16` → IMP-16 `state=open`, label `needs-codex-review`.
+
+Stage-axis cross-references:
+
+- Stage 1 exit report (binding): "IMP-07 = pending"; "U2 hard_link_to_imp07: yes"; "U2 gate: blocked_until_imp07_lands" — assumption now superseded by IMP-07 actual closure.
+- Stage 2 exit report (binding): u1~u11 plan; Codex r3 verdict ok per unit.
+- Stage 3 exit report (binding): 12 untracked artifacts created; Codex #26 r11 PASS for u11.
+- Stage 4 exit report (binding): targeted 56 PASS / regression 15 PASS / py_compile clean / no `src.content_verifier` import / no IMP-07 forbidden-coupling tokens in `src/`.
+- Stage 5 exit report (binding): commit `23ba8b6` on both remotes; `git diff --cached --stat` empty post-commit; Codex #28 PASS.
+
+rewind_target: n/a
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_final-close_codex_r1.md

@@ -0,0 +1,62 @@
+[Codex #29] Stage 6 final-close Round #1 independent verification for IMP-16.
+
+Decision: PASS / close OK, with dependency note.
+
+I verified the Stage 5 commit, remote state, committed file scope, deterministic guardrails, tests, and the IMP-07 hard-link state. The implementation matches the binding IMP-16 contract: U1 landed as a Phase Z-owned deterministic verification utility, and U2 landed only as a design contract with no runtime Step 1/2/14/21/22 wiring.
+
+Important reconciliation: IMP-07 is not pending now. Gitea issue #7 is closed at 2026-05-15T18:28:24+09:00, and the closing comments state the reverse path is not needed because manual HTML edits remain HTML-only. That means IMP-16-U2 has no current production trigger. This does not require a rewind because the committed IMP-16 code did not implement U2 wiring, did not add reverse-path runtime hooks, and preserved the hard gate. If a future reverse-path issue reopens the product direction, U2 should be revalidated against that new parent rather than assuming IMP-07 will land.
+
+=== EVIDENCE ===
+Commands run:
+- `git rev-parse HEAD`
+- `git status --short`
+- `git show --stat --oneline --no-renames HEAD`
+- `git show --name-only --format= --no-renames HEAD`
+- `git ls-remote origin refs/heads/main`
+- `git ls-remote slide2 refs/heads/main`
+- `git show --format= --no-renames HEAD -- src/phase_z2_verification_utils.py | Select-String -Pattern '^\\+import |^\\+from '`
+- `rg -n "html_to_slide_mdx|edited_html_to_mdx|reverse_path|reentry_" src`
+- `rg -n "src\\.content_verifier|content_verifier|FORBIDDEN_KEI_MEMOS|generate_with_retry|REQUIRED_PATTERNS|verify_structure|verify_area|verify_all_areas|httpx|SSE|html_to_slide_mdx|edited_html_to_mdx|reverse_path|reentry_" src docs/architecture/IMP-16-U2-WIRING-DESIGN.md tests/phase_z2`
+- `pytest -q tests/phase_z2/test_pz2_vu_*.py` (PowerShell glob did not expand; rerun below)
+- `pytest -q @(Get-ChildItem -LiteralPath 'tests\\phase_z2' -Filter 'test_pz2_vu_*.py' | ForEach-Object { $_.FullName })`
+- `pytest -q tests/orchestrator_unit/test_subprocess_cleanup.py`
+- Gitea API: `GET /repos/Kyeongmin/C.E.L_Slide_test2/issues/7`
+- Gitea API: `GET /repos/Kyeongmin/C.E.L_Slide_test2/issues/7/comments?limit=50`
+- Gitea API: `GET /repos/Kyeongmin/C.E.L_Slide_test2/issues/16`
+
+Files checked:
+- `src/phase_z2_verification_utils.py`
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`
+- `tests/phase_z2/test_pz2_vu_dataclass.py`
+- `tests/phase_z2/test_pz2_vu_extract.py`
+- `tests/phase_z2/test_pz2_vu_normalize.py`
+- `tests/phase_z2/test_pz2_vu_keywords.py`
+- `tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `tests/phase_z2/test_pz2_vu_sentence_split.py`
+- `tests/phase_z2/test_pz2_vu_match_helper.py`
+- `tests/phase_z2/test_pz2_vu_preservation.py`
+- `tests/phase_z2/test_pz2_vu_invented.py`
+- `tests/phase_z2/test_pz2_vu_integration.py`
+
+Test results:
+- IMP-16 targeted tests: 56 passed in 0.18s.
+- Regression guard: 15 passed in 7.59s.
+
+Commit SHA:
+- Local HEAD: `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`
+- `origin/main`: `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`
+- `slide2/main`: `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`
+
+Verified facts:
+- HEAD commit message is `feat(IMP-16): U1 H3 verification utility port + U2 wiring design`.
+- HEAD contains exactly the 12 IMP-16 files listed in the Stage 5 exit report.
+- The commit stat is 12 files changed, 1156 insertions, 0 deletions.
+- `src/phase_z2_verification_utils.py` imports only `__future__`, `re`, `dataclasses`, `difflib`, and `html.parser`.
+- No runtime `src/` implementation hit exists for `html_to_slide_mdx`, `edited_html_to_mdx`, `reverse_path`, or `reentry_`.
+- IMP-16-U2 is documentation-only. Runtime wiring is still absent, as required by the binding gate.
+- References to `src.content_verifier`, `REQUIRED_PATTERNS`, `verify_structure`, `verify_area`, `verify_all_areas`, `FORBIDDEN_KEI_MEMOS`, `generate_with_retry`, `httpx`, and `SSE` in the IMP-16 commit are either U2 documentation or test/comment reference text; the new runtime module does not import or call them.
+- Existing dirty working-tree files and unrelated untracked artifacts remain outside the IMP-16 commit scope.
+- Gitea issue #16 is still open at verification time.
+- Gitea issue #7 is closed, and its closing comments explicitly record the product decision that edited HTML is not converted back to MDX in the current direction.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_problem-review_claude_r1.md

@@ -0,0 +1,84 @@
+[Claude #1] Stage 1 — IMP-16 problem review (B-2 verification 보조 axis)
+
+### Root-cause framing
+
+IMP-16 is the verification-helper sibling of IMP-07 (B-2 main: edited HTML → MDX → pipeline re-entry). When IMP-07 lands, the re-entered MDX needs deterministic checks on three axes — text preserved, structure intact, trace surfaced — without re-introducing AI/Kei into the normal path. The Phase Q file `src/content_verifier.py` already contains the relevant deterministic utilities, but it is classified **Reference Only** (PHASE-Q-AUDIT.md §2.7, H3) — direct import is forbidden because the same file also hosts Archive-only Kei assets (`FORBIDDEN_KEI_MEMOS`, `generate_with_retry`). IMP-16 scope is therefore to **port the H3 deterministic subset into a new Phase Z module**, NOT to import `content_verifier.py` from Phase Z pipeline. The port is splittable into two units; only the wiring-free unit can land now under the `hard link: IMP-07` gate.
+
+### Verified facts (value + path + upstream)
+
+- IMP-16 backlog row = `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67` — priority ↓ low, hard link IMP-07, guardrail "AI/Kei verification 회귀 X / utility deterministic"
+- IMP-07 backlog row = same doc line 51 — status `pending`. `git log --all --oneline | grep -iE "IMP-07|html_to_slide_mdx|B-2 reverse|reverse path"` → 0 hits. `grep -rn "html_to_slide_mdx\|reverse_path\|edited_html_to_mdx" --include="*.py"` → 0 hits. **IMP-07 (B-2 main) is unimplemented in repo.**
+- Source insight = `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:122` (§3 axis registry "B-2 verification 보조") + `docs/architecture/PHASE-Q-AUDIT.md:766` (§2.7 H3 footnote)
+- H3 Reference Only assets in `src/content_verifier.py`:
+  - `VerificationResult` dataclass (line 29)
+  - `_TextExtractor` (43) + `extract_text_from_html` (66)
+  - `normalize_for_comparison` (97) + `_PARTICLES` (78) + `_ENDING_NORMALIZE` (85)
+  - `extract_keywords` (120), `strip_meta_lines` (150), `split_into_sentences` (174)
+  - `verify_text_preservation` L1 (206), `detect_invented_text` (276)
+  - `verify_structure` L3 (395) + `REQUIRED_PATTERNS` (382)
+- Archive assets in same file (**EXCLUDED** from IMP-16): `verify_no_forbidden_content` (338), `FORBIDDEN_KEI_MEMOS` (322), `FORBIDDEN_LABELS_IN_KEYMSG` (331), `generate_with_retry` (async) — H4/H5 markers in §2.7.
+- Phase Z import isolation (verified now): `grep -n "extract_text_from_html\|normalize_for_comparison\|extract_keywords\|verify_text_preservation" src/phase_z2_pipeline.py` → 0 hits. Phase Q `content_verifier.py` is currently imported only by `src/pipeline.py`, `scripts/run_from_artifacts.py`, `scripts/test_phase_s.py`, `scripts/test_phase_t_full.py`. **No existing Phase Z coupling to refactor.**
+- `REQUIRED_PATTERNS` (Phase Q line 382) uses `body_bg/core/sidebar/footer` area abstraction — explicitly mismatched with Phase Z unit/zone/preset/frame contract abstraction (PHASE-Q-AUDIT.md §2.7 H2/H3 footnotes). Values cannot be reused; pattern only. L3 port belongs to IMP-20 (H2), not IMP-16.
+- Related steps cross-ref (PHASE-Q-INSIGHT-TO-22STEP-MAP.md:100/107): Step 14 (visual_check, partial) + Step 21 (debug.json trace, partial) were just extended by IMP-15 (commits e9b3d2e/2827622/535c484/614c533 — `image_aspect_mismatch` + `table_self_overflow` + classifier + debug.json surfacing). Step 1/2 = MDX normalize + section parse; Step 22 = UI export. **IMP-16 sits on top of these steps as a separate verification layer, not inside any of them.** IMP-15 = visual axis; IMP-16 = text axis. Non-overlapping.
+- HEAD = `614c533`; branch = `main`; working tree dirty on orchestrator infra (`orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py`, `.orchestrator/`, `chromedriver/`, …) — all unrelated WIP, MUST NOT be staged under this issue.
+
+### Scope-lock
+
+The `hard link: IMP-07` clause means IMP-16 cannot **wire** into Phase Z pipeline until IMP-07's reverse-path output schema exists. But the H3 subset port can be split into two units; only Unit A is implementable now.
+
+**SCOPE-LOCKED (Unit A — implementable in this cycle):** deterministic text-axis utility port into a new module `src/phase_z2_verification.py`. Pure utility code, zero pipeline wiring.
+
+Ported subset (text axis only):
+- `VerificationResult` (dataclass — Phase Z trace schema additive shape)
+- `extract_text_from_html` + `_TextExtractor`
+- `normalize_for_comparison` + `_PARTICLES` + `_ENDING_NORMALIZE`
+- `extract_keywords` + `strip_meta_lines` + `split_into_sentences`
+- `verify_text_preservation` (L1) + `detect_invented_text`
+
+**OUT OF SCOPE (Unit B — blocked by IMP-07):** pipeline wiring at Step 14 / Step 21 / Step 22. No consumer of the ported utilities lands until IMP-07 produces the `(original_mdx, edited_html_text)` pair.
+
+**OUT OF SCOPE (axis bleed — belongs to IMP-20):** `verify_structure` (L3) + `REQUIRED_PATTERNS` port. Phase Z requires per-frame-contract pattern dict, not Phase Q area abstraction — H2 reference, separate axis.
+
+**OUT OF SCOPE (Archive H4/H5):** `verify_no_forbidden_content` + `FORBIDDEN_KEI_MEMOS` + `FORBIDDEN_LABELS_IN_KEYMSG` + `generate_with_retry`. Kei-coupled.
+
+**OUT OF SCOPE (Phase Q file mutation):** `src/content_verifier.py` is untouched. Phase Q pipeline keeps it; IMP-16 is copy-and-adapt-into-new-module, not move/mutate. G3 isolation invariant.
+
+### Guardrails
+
+- G1 (PZ-1, RULE 0): AI=0 normal path — port subset is deterministic only; H4/H5 Archive assets EXCLUDED.
+- G2 (RULE 7): no MDX 03/04/05 hardcoding — utilities are content-agnostic by construction (text/regex/HTML parse only). Tests must use synthetic fixtures, not real MDX samples.
+- G3: `src/content_verifier.py` (Phase Q) remains untouched. Phase Q pipeline (`src/pipeline.py` + 3 scripts) keeps its import. Port is *copy + adapt into new Phase Z module*.
+- G4 (H3 footnote): Phase Z `REQUIRED_PATTERNS`-equivalent EXCLUDED from this cycle — belongs to IMP-20 (H2). No L3 port in Unit A.
+- G5 (hard-link IMP-07): NO pipeline wiring in this cycle. Ported utilities are unused until IMP-07 lands. Stage 4 verification = unit tests only; no Step 14/21/22 integration.
+- G6: tests deterministic only — fixtures + pure assertions; no Anthropic / Kei client invocation; no Selenium (visual_check belongs to IMP-15 / IMP-01).
+- G7 (RULE 13 anchor sync): in commit-push stage, `docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` (if it tracks IMP-16) + IMP-16 backlog row status field update in same commit (anchor SHA discipline). INSIGHT-MAP §3 row reference unchanged (still "pending" until Unit B lands).
+- G8 (PZ-2): 1 turn = 1 step — Stage 1 closes here; no jumping ahead to plan. No mention of Unit B implementation timing.
+- G9 (Workflow atomicity): 1 commit = 1 decision unit = Unit A port only. Existing working-tree WIP (orchestrator.py etc.) MUST NOT be staged under this issue.
+
+### Risk
+
+Low. Pure utility port + tests. Failure modes:
+- (a) Test fixture choice accidentally embeds MDX 03/04/05 content → RULE 7 violation; rewind to `stage_3_edit` with synthetic-fixture rewrite.
+- (b) Port surface accidentally pulls H4/H5 symbols → G1 violation; rewind to `stage_3_edit`.
+- (c) New module accidentally imported by `phase_z2_pipeline.py` → G5 violation (premature wiring); rewind to `stage_3_edit`.
+- (d) Phase Q `content_verifier.py` modified by accident → G3 violation; rewind to `stage_3_edit`.
+
+### Open questions for Codex r1
+
+1. **Module name confirm**: `src/phase_z2_verification.py` proposed. Alternative: `src/phase_z2_text_verification.py` (axis-prefixed, leaves room for `src/phase_z2_visual_verification.py` if IMP-01/IMP-15-extension ever ports there). Recommend axis-prefixed for naming hygiene; defer to Codex.
+2. **`verify_text_preservation` + `detect_invented_text` port boundary**: these depend on `normalize_for_comparison` + `extract_keywords` — full transitive closure ports cleanly. Confirm full transitive closure (proposed) vs minimal extract-only (text utilities, defer L1 checks to Unit B).
+3. **Unit-test scope**: parity with `src/content_verifier.py`'s existing test surface (if any) vs Phase-Z-style fixture tests. Recommend Phase-Z-style with explicit RULE 7 guard.
+
+### Stage 1 lock summary
+
+- Root cause: B-2 reverse path (IMP-07) needs a deterministic verification layer; H3 utilities in Phase Q `content_verifier.py` are Reference Only and cannot be imported (Archive coupling + abstraction mismatch). Phase Z has no text-axis verification module today.
+- Scope locked: **Unit A only** — port H3 deterministic text-axis subset into new module `src/phase_z2_verification.py` (or `src/phase_z2_text_verification.py` — naming TBD); NO pipeline wiring; NO L3; NO Archive symbols; tests deterministic+synthetic.
+- Unit B (wiring) explicitly deferred per `hard link: IMP-07`.
+- Guardrails G1–G9 above.
+
+=== EVIDENCE ===
+- Commands: `git log --oneline -20`; `git log --all --oneline | grep -iE "IMP-07|html_to_slide_mdx|B-2 reverse|reverse path"` (0 hits); `grep -rn "html_to_slide_mdx\|reverse_path\|edited_html_to_mdx" --include="*.py"` (0 hits — IMP-07 unimplemented); `grep -rln "from src.content_verifier\|import content_verifier" --include="*.py"` → 4 hits, all Phase Q (`scripts/run_from_artifacts.py`, `scripts/test_phase_s.py`, `scripts/test_phase_t_full.py`, `src/pipeline.py`); `grep -n "extract_text_from_html\|normalize_for_comparison\|extract_keywords\|verify_text_preservation" src/phase_z2_pipeline.py` (0 hits); `grep -n "^def \|^class \|^_PARTICLES\|^_ENDING_NORMALIZE\|^FORBIDDEN\|^REQUIRED_PATTERNS" src/content_verifier.py` confirms symbol locations 29/43/66/78/85/97/120/150/174/206/276/322/331/338/382/395/455/490; `git rev-parse HEAD` → `614c53358ecaa92e9f6fc979b2399da028e15d41`; `git status --porcelain` → orchestrator WIP only, no `src/phase_z2_*` or `src/content_verifier*` dirty.
+- Files read: `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (IMP-07 line 51, IMP-16 line 67, hard-link convention line 30); `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` lines 90–140 (§2-B/§2-C/§3 axis registry); `docs/architecture/PHASE-Q-AUDIT.md` lines 100–110 (§1 lens B-2 row) + 689–780 (§2.7 content_verifier audit, H3 footnote at line 766, H4/H5 Archive markers 768/770); `src/content_verifier.py` symbol map; `.orchestrator/drafts/49_stage_problem-review_claude_r1.md` (precedent format check — Claude #1 problem-review structure).
+- Facts cross-checked: IMP-07 (B-2 main) unimplemented in repo → Unit B must defer (G5); Phase Z pipeline does not currently import `content_verifier` → no refactor risk; `REQUIRED_PATTERNS` Phase Q area ≠ Phase Z frame-contract abstraction (audit §2.7 H2/H3 explicit) → L3 port out of scope (IMP-20); IMP-15 (Step 14 visual_check) just landed in 4 commits but is visual axis, IMP-16 is text axis — non-overlapping.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_problem-review_codex_r1.md

@@ -0,0 +1,68 @@
+[Codex #1] Stage 1 problem review for IMP-16
+
+Root cause is verified. IMP-16 is a verification-helper axis for the future B-2 edited-HTML-to-MDX reverse path, not the main reverse path itself. The only concrete source assets found today are the deterministic/reference utilities in `src/content_verifier.py`; Phase Q audit classifies those utilities as Reference Only for B-2 verification use, while keeping Kei/AI retry and forbidden-content assets as Archive Candidate. I did not find an existing Phase Z implementation of the B-2 reverse path or a Phase Z import/coupling that needs refactor.
+
+Scope lock:
+- IMP-16 must not import `src.content_verifier` directly from Phase Z. The source file mixes deterministic utilities with Archive-only Kei/AI assets, so the acceptable future implementation shape is a Phase Z-owned deterministic helper, using H3 as reference only.
+- IMP-16 must not implement the main edited HTML -> MDX reverse path. That remains IMP-07.
+- Because the backlog defines `hard link: IMP-07` and IMP-07 is still `pending` with no implementation hits, any runtime wiring into Step 1/2/14/21/22 should stay blocked until IMP-07 exists. A planning-only or isolated utility design can be described, but implementation must not pretend the B-2 main path exists.
+- `REQUIRED_PATTERNS` values from `src/content_verifier.py` must not be reused as Phase Z contract checks. The audit explicitly says Phase Q areas are `body_bg/core/sidebar/footer`, while Phase Z uses unit/zone/preset/frame-contract abstractions. The reusable part is the pattern-checking approach, not the constants.
+
+Missing files / wrong assumptions flagged:
+- No `html_to_slide_mdx`, `edited_html_to_mdx`, or B-2 reverse implementation was found in Python/JS/TS/MD docs outside architecture references.
+- No existing Phase Z B-2 verification helper module was found. `src/fit_verifier.py` exists, but it is Phase V fit/height verification and not a B-2 reverse-verification helper.
+- The statement "wiring-free unit can land now under the hard-link gate" should be treated narrowly. It is only safe as future-plan decomposition or an isolated deterministic module proposal; it is not safe as Step 1/2/14/21/22 pipeline wiring before IMP-07 lands.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\.orchestrator\tmp\codex_prompt.txt'`
+- `rg -n "IMP-16|IMP-07|B-2 verification|hard link|AI/Kei verification|utility deterministic" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs/architecture/PHASE-Q-AUDIT.md docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md`
+- `rg -n "class VerificationResult|@dataclass|class _TextExtractor|def extract_text_from_html|def normalize_for_comparison|_PARTICLES|_ENDING_NORMALIZE|def extract_keywords|def strip_meta_lines|def split_into_sentences|def verify_text_preservation|def detect_invented_text|REQUIRED_PATTERNS|def verify_structure|FORBIDDEN_KEI_MEMOS|FORBIDDEN_LABELS_IN_KEYMSG|def verify_no_forbidden_content|def generate_with_retry" src/content_verifier.py`
+- `rg -n "content_verifier|extract_text_from_html|normalize_for_comparison|extract_keywords|verify_text_preservation|detect_invented_text|verify_structure|REQUIRED_PATTERNS" src scripts tests docs -g "*.py" -g "*.md"`
+- `git log --all --oneline --decorate --grep="IMP-07\|html_to_slide_mdx\|B-2 reverse\|reverse path"`
+- `rg -n "html_to_slide_mdx|reverse_path|edited_html_to_mdx|Edited HTML|B-2 Edited|B-2 reverse|reverse path" . -g "*.py" -g "*.md" -g "*.ts" -g "*.tsx" -g "*.js" -g "*.jsx"`
+- `rg --files src | rg "phase_z|z2"`
+- `rg -n "content_verifier|extract_text_from_html|normalize_for_comparison|extract_keywords|verify_text_preservation|detect_invented_text|verify_structure|REQUIRED_PATTERNS" src -g "*phase_z*.py" -g "*z2*.py"`
+- `rg -n "from src\.content_verifier|import src\.content_verifier|content_verifier import|import content_verifier" src scripts tests -g "*.py"`
+- `rg -n "Step 1|Step 2|Step 14|Step 21|Step 22" docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md`
+- `rg -n "§3|B-2 verification|Step 1|Step 2|Step 14|Step 21|Step 22" docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `rg --files src tests | rg -i "verif|reverse|html.*mdx|mdx.*html|phase_z.*check|trace"`
+- `rg -n "VerificationResult|text_preservation|invented|extract_text_from_html|normalize_for_comparison|B-2 verification|verification helper|reverse verification" src tests docs -g "*.py" -g "*.md"`
+- `Get-Content -LiteralPath 'src\fit_verifier.py' | Select-Object -First 220`
+- `rg -n "fit_verifier|verify_fit|overflow|visual_check|B-2|reverse|MDX|content_verifier" src tests -g "*.py"`
+- `git status --short --branch`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-AUDIT.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_router.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_retry.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_placement_planner.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_mapper.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_internal_region_planner.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_failure_router.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_content_extractor.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_composition.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_classifier.py`
+- `D:\ad-hoc\kei\design_agent\src\fit_verifier.py`
+
+Test results: Not run. Current Stage 1 task explicitly says "NO pytest"; only grep/find verification was performed.
+
+Verified facts:
+- `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67` defines IMP-16 as low priority, Step 1/2/14/21/22, `§3 H3 Reference Only`, B-2 verification helper layer, deterministic utility guardrail, and `hard link: IMP-07`.
+- `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:51` defines IMP-07 as the B-2 Edited HTML -> MDX reverse path and its status is `pending`.
+- `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:30` defines hard links as requiring the predecessor issue before proceeding.
+- `PHASE-Q-INSIGHT-TO-22STEP-MAP.md:122` registers `B-2 verification 보조` for Step 1, 2, 14, 21, 22 and points it to `§2.7 H3`.
+- `PHASE-Q-AUDIT.md:733-742` classifies the relevant `content_verifier.py` deterministic assets as Reference Only.
+- `PHASE-Q-AUDIT.md:764-766` says `verify_structure` pattern is only reference and `REQUIRED_PATTERNS` values are Phase Q area-specific, while H3 says B-2 can reference text extraction/normalization/comparison utilities but direct integration is disallowed.
+- `PHASE-Q-AUDIT.md:751` and `PHASE-Q-AUDIT.md:780` say `content_verifier.py` has Salvage Candidate 0 and direct integration is not allowed.
+- `src/content_verifier.py` contains the claimed deterministic utilities at the expected locations: `VerificationResult`, `_TextExtractor`, `extract_text_from_html`, normalization helpers, keyword/sentence helpers, `verify_text_preservation`, `detect_invented_text`, `REQUIRED_PATTERNS`, and `verify_structure`.
+- `src/content_verifier.py` also contains excluded Archive/Kei/AI assets: `FORBIDDEN_KEI_MEMOS`, `FORBIDDEN_LABELS_IN_KEYMSG`, `verify_no_forbidden_content`, and async `generate_with_retry`.
+- Current imports of `src.content_verifier` are limited to `src/pipeline.py`, `scripts/run_from_artifacts.py`, `scripts/test_phase_s.py`, and `scripts/test_phase_t_full.py`; no Phase Z module import was found.
+- No B-2 reverse implementation was found by searches for `html_to_slide_mdx`, `reverse_path`, `edited_html_to_mdx`, `B-2 reverse`, or related terms. Hits are architecture docs only.
+- The working tree is dirty before this review; I did not touch existing code or run tests.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_simulation-plan_claude_r1.md

@@ -0,0 +1,81 @@
+[Claude #2] Stage 2 — IMP-16 simulation plan r1. Unit-1 only (H3 deterministic utility port). U2 wiring (Step 1/2/14/21/22) deferred — hard-link gate on IMP-07.
+
+=== IMPLEMENTATION_UNITS ===
+```yaml
+- id: u1
+  summary: New src/phase_z2_verification_utils.py — docstring, imports, VerificationResult dataclass.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_dataclass.py]
+  tests: [tests/phase_z2/test_pz2_vu_dataclass.py]
+  estimate_lines: 30
+- id: u2
+  summary: Port _TextExtractor + extract_text_from_html (HTMLParser; style/script skip).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_extract.py]
+  tests: [tests/phase_z2/test_pz2_vu_extract.py]
+  estimate_lines: 30
+- id: u3
+  summary: Port _PARTICLES, _ENDING_NORMALIZE, normalize_for_comparison.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_normalize.py]
+  tests: [tests/phase_z2/test_pz2_vu_normalize.py]
+  estimate_lines: 40
+- id: u4
+  summary: Port extract_keywords (>=3-char + trailing-particle strip; uses u3).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_keywords.py]
+  tests: [tests/phase_z2/test_pz2_vu_keywords.py]
+  estimate_lines: 20
+- id: u5
+  summary: Port _META_PREFIXES + strip_meta_lines.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_meta_strip.py]
+  tests: [tests/phase_z2/test_pz2_vu_meta_strip.py]
+  estimate_lines: 32
+- id: u6
+  summary: Port split_into_sentences (calls u5).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_sentence_split.py]
+  tests: [tests/phase_z2/test_pz2_vu_sentence_split.py]
+  estimate_lines: 28
+- id: u7
+  summary: Private _match_sentence_against_html (kw-ratio + SequenceMatcher); keeps u8 atomic.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_match_helper.py]
+  tests: [tests/phase_z2/test_pz2_vu_match_helper.py]
+  estimate_lines: 30
+- id: u8
+  summary: Port verify_text_preservation (composes u3/u4/u6/u7; threshold preserved).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_preservation.py]
+  tests: [tests/phase_z2/test_pz2_vu_preservation.py]
+  estimate_lines: 40
+- id: u9
+  summary: Port detect_invented_text (kw-ratio<0.4; min_length=15; CSS-value skip).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_invented.py]
+  tests: [tests/phase_z2/test_pz2_vu_invented.py]
+  estimate_lines: 42
+- id: u10
+  summary: E2E smoke on samples/mdx_batch/03.mdx + synth HTML — no pipeline import.
+  files: [tests/phase_z2/test_pz2_vu_integration.py]
+  tests: [tests/phase_z2/test_pz2_vu_integration.py]
+  estimate_lines: 40
+```
+
+### Rationale (≤3 lines per unit)
+- u1: anchor; later units extend same flat module; zero content_verifier imports (Stage 1 lock).
+- u2/u3/u4/u6: stdlib + regex; pure deterministic.
+- u5: _META_PREFIXES = Kei analyzer-instruction filter (제목 라벨:, 표현 의도:, …) — orthogonal to Phase Q REQUIRED_PATTERNS (NOT ported).
+- u7: helper extracted so u8 ≤50 lines.
+- u8/u9: compose earlier units; threshold/ratio knobs preserved.
+- u10: composition sanity — no phase_z2_pipeline.py import, no AI.
+
+### Out-of-scope (deferred / blocked)
+- U2 wiring (Step 1 re-entered MDX validation / Step 2 text preservation / Step 14 structure / Step 21 trace / Step 22 report) — hard-link blocked until IMP-07 lands. No phase_z2_pipeline.py edits.
+- REQUIRED_PATTERNS L382 / verify_structure L395 / verify_area L455 / verify_all_areas L490 — Phase Q area-specific values; U2-time redesign vs Phase Z frame-contract abstraction. NOT ported.
+- FORBIDDEN_KEI_MEMOS / verify_no_forbidden_content (H4). generate_with_retry (H5). Selenium/Vision L4/L5 (separate axis).
+- No `from src.content_verifier import …` in any phase_z2_* module.
+
+### Rollback
+Each unit additive to a NEW file with zero existing call sites — rollback = `git revert <unit_commit>`. No pipeline regression surface until U2 lands.
+
+=== EVIDENCE ===
+- src/content_verifier.py port-target lines re-verified: VerificationResult L28-36; _TextExtractor L43-63; extract_text_from_html L66-70; _PARTICLES L78-82; _ENDING_NORMALIZE L85-94; normalize_for_comparison L97-117; extract_keywords L120-134; _META_PREFIXES L138-147; strip_meta_lines L150-171; split_into_sentences L174-199; verify_text_preservation L206-273; detect_invented_text L276-315.
+- src/ flat naming verified — new module = `src/phase_z2_verification_utils.py` (not subpackage). tests/phase_z2/ exists.
+- 0 hits for `phase_z2_verification` / `verification_utils` in repo.
+- 0 hits for `content_verifier` imported by src/phase_z2_*.py (PHASE-Q-AUDIT §2.7 H3 Reference Only preserved).
+- IMP-07 still pending (Stage 1 grep html_to_slide_mdx|edited_html_to_mdx|reverse_path → 0 hits) → U2 units intentionally absent.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_simulation-plan_claude_r2.md

@@ -0,0 +1,93 @@
+[Claude #3] Stage 2 r2 — IMP-16. Addresses codex r1 `missing_unit` via u11 (IMP-16-U2 design-only md, no pipeline code). u1-u10 unchanged (codex r1: ok).
+
+=== IMPLEMENTATION_UNITS ===
+```yaml
+- id: u1
+  summary: New src/phase_z2_verification_utils.py — docstring, imports, VerificationResult dataclass.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_dataclass.py]
+  tests: [tests/phase_z2/test_pz2_vu_dataclass.py]
+  estimate_lines: 30
+- id: u2
+  summary: Port _TextExtractor + extract_text_from_html (HTMLParser; style/script skip).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_extract.py]
+  tests: [tests/phase_z2/test_pz2_vu_extract.py]
+  estimate_lines: 30
+- id: u3
+  summary: Port _PARTICLES, _ENDING_NORMALIZE, normalize_for_comparison.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_normalize.py]
+  tests: [tests/phase_z2/test_pz2_vu_normalize.py]
+  estimate_lines: 40
+- id: u4
+  summary: Port extract_keywords (>=3-char + trailing-particle strip; uses u3).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_keywords.py]
+  tests: [tests/phase_z2/test_pz2_vu_keywords.py]
+  estimate_lines: 20
+- id: u5
+  summary: Port _META_PREFIXES + strip_meta_lines.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_meta_strip.py]
+  tests: [tests/phase_z2/test_pz2_vu_meta_strip.py]
+  estimate_lines: 32
+- id: u6
+  summary: Port split_into_sentences (calls u5).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_sentence_split.py]
+  tests: [tests/phase_z2/test_pz2_vu_sentence_split.py]
+  estimate_lines: 28
+- id: u7
+  summary: Private _match_sentence_against_html (kw-ratio + SequenceMatcher); keeps u8 atomic.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_match_helper.py]
+  tests: [tests/phase_z2/test_pz2_vu_match_helper.py]
+  estimate_lines: 30
+- id: u8
+  summary: Port verify_text_preservation (composes u3/u4/u6/u7; threshold 0.7 preserved).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_preservation.py]
+  tests: [tests/phase_z2/test_pz2_vu_preservation.py]
+  estimate_lines: 40
+- id: u9
+  summary: Port detect_invented_text (kw-ratio<0.4; min_length=15; CSS-value skip).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_invented.py]
+  tests: [tests/phase_z2/test_pz2_vu_invented.py]
+  estimate_lines: 42
+- id: u10
+  summary: E2E smoke on samples/mdx_batch/03.mdx + synth HTML — no pipeline import.
+  files: [tests/phase_z2/test_pz2_vu_integration.py]
+  tests: [tests/phase_z2/test_pz2_vu_integration.py]
+  estimate_lines: 40
+- id: u11
+  summary: Design-only md — IMP-16-U2 wiring contract (Step 1/2/14/21/22 hook points + frame-contract pattern dict redesign + IMP-07 gate). NO Python.
+  files: [docs/architecture/IMP-16-U2-WIRING-DESIGN.md]
+  tests: []
+  estimate_lines: 50
+```
+
+### Rationale
+- u1: anchor; subsequent units extend same flat module; zero `src.content_verifier` imports.
+- u2/u3/u4/u6: stdlib HTMLParser + regex; pure deterministic; locale rules verbatim.
+- u5: `_META_PREFIXES` = Kei analyzer-instruction filter; orthogonal to Phase Q `REQUIRED_PATTERNS` (not ported).
+- u7: helper extracted so u8 stays ≤50 lines (kw-ratio gate + SequenceMatcher fuzzy).
+- u8: composes u3/u4/u6/u7; threshold from `src/content_verifier.py:206-274` verbatim.
+- u9: pure regex/string; ported verbatim from source.
+- u10: smoke on real sample + synth HTML; NO `src.phase_z2_pipeline` import.
+- u11: md doc captures U2 wiring contract — hook points (Step 1/2/14/21/22), frame-contract pattern dict redesign principles (no Phase Q area-name reuse), IMP-07 hard-link gate. Addresses codex r1 `missing_unit` without violating Stage 1 scope_lock.
+
+### Out-of-scope
+- B-2 main reverse path (IMP-07).
+- Direct `src/content_verifier.py` import from any `phase_z2_*` module.
+- `REQUIRED_PATTERNS` value reuse.
+- `FORBIDDEN_KEI_MEMOS` / `generate_with_retry` (H4/H5 archive).
+- Step 14 image_aspect/table overflow (IMP-15).
+- u11 implementation (md only — no `.py`, no test, no wiring).
+- IMP-16-U2 code wiring (blocked behind IMP-07).
+
+### Rollback
+Per-unit revert. u1-u10 = append-only to one new file (no cross-file edits). u11 = doc-only, zero code impact.
+
+=== EVIDENCE ===
+- Stage 1 exit report (commit 614c533): IMP-16-U1 = free_to_land; IMP-16-U2 = blocked_until_imp07_lands, hard_link_to_imp07: yes.
+- Codex r1 verdict u1-u10 = ok (each ≤50 lines AND ≤3 files).
+- Codex r1 `missing_unit: IMP-16-U2` → addressed by u11 (design-only md, no pipeline code, no IMP-07 gate violation).
+- `src/content_verifier.py` symbols verified Stage 1 (lines 29-490).
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67`: IMP-16 priority ↓ low, hard link IMP-07.
+- Repo grep `html_to_slide_mdx|edited_html_to_mdx|reverse_path` → 0 hits (IMP-07 pending).
+- u11 path new file, no overwrite risk.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_simulation-plan_claude_r3.md

@@ -0,0 +1,93 @@
+[Claude #4] Stage 2 r3 — IMP-16. Plan converged r2 (Codex r2: u1-u11 ok). r3 = no-change repost. U1 ports H3 utility; U2 wiring deferred behind IMP-07 gate via u11 design-only md.
+
+=== IMPLEMENTATION_UNITS ===
+```yaml
+- id: u1
+  summary: New src/phase_z2_verification_utils.py — docstring, imports, VerificationResult dataclass.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_dataclass.py]
+  tests: [tests/phase_z2/test_pz2_vu_dataclass.py]
+  estimate_lines: 30
+- id: u2
+  summary: Port _TextExtractor + extract_text_from_html (HTMLParser; style/script skip).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_extract.py]
+  tests: [tests/phase_z2/test_pz2_vu_extract.py]
+  estimate_lines: 30
+- id: u3
+  summary: Port _PARTICLES, _ENDING_NORMALIZE, normalize_for_comparison.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_normalize.py]
+  tests: [tests/phase_z2/test_pz2_vu_normalize.py]
+  estimate_lines: 40
+- id: u4
+  summary: Port extract_keywords (>=3-char + trailing-particle strip; uses u3).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_keywords.py]
+  tests: [tests/phase_z2/test_pz2_vu_keywords.py]
+  estimate_lines: 20
+- id: u5
+  summary: Port _META_PREFIXES + strip_meta_lines.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_meta_strip.py]
+  tests: [tests/phase_z2/test_pz2_vu_meta_strip.py]
+  estimate_lines: 32
+- id: u6
+  summary: Port split_into_sentences (calls u5).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_sentence_split.py]
+  tests: [tests/phase_z2/test_pz2_vu_sentence_split.py]
+  estimate_lines: 28
+- id: u7
+  summary: Private _match_sentence_against_html (kw-ratio + SequenceMatcher); keeps u8 atomic.
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_match_helper.py]
+  tests: [tests/phase_z2/test_pz2_vu_match_helper.py]
+  estimate_lines: 30
+- id: u8
+  summary: Port verify_text_preservation (composes u3/u4/u6/u7; threshold 0.7 preserved).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_preservation.py]
+  tests: [tests/phase_z2/test_pz2_vu_preservation.py]
+  estimate_lines: 40
+- id: u9
+  summary: Port detect_invented_text (kw-ratio<0.4; min_length=15; CSS-value skip).
+  files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_invented.py]
+  tests: [tests/phase_z2/test_pz2_vu_invented.py]
+  estimate_lines: 42
+- id: u10
+  summary: E2E smoke on samples/mdx_batch/03.mdx + synth HTML — no pipeline import.
+  files: [tests/phase_z2/test_pz2_vu_integration.py]
+  tests: [tests/phase_z2/test_pz2_vu_integration.py]
+  estimate_lines: 40
+- id: u11
+  summary: Design-only md — IMP-16-U2 wiring contract (Step 1/2/14/21/22 hook points + frame-contract pattern dict redesign + IMP-07 gate). NO Python.
+  files: [docs/architecture/IMP-16-U2-WIRING-DESIGN.md]
+  tests: []
+  estimate_lines: 50
+```
+
+### Rationale
+- u1: anchor; later units extend same flat module; zero `src.content_verifier` imports.
+- u2/u3/u4/u6: stdlib HTMLParser + regex; pure deterministic; Korean locale rules verbatim.
+- u5: `_META_PREFIXES` = Kei analyzer-instruction filter; orthogonal to Phase Q `REQUIRED_PATTERNS` (not ported).
+- u7: helper extracted so u8 stays ≤50 lines (kw-ratio gate + SequenceMatcher fuzzy).
+- u8: composes u3/u4/u6/u7; threshold from `src/content_verifier.py:206-274` verbatim.
+- u9: pure regex/string; verbatim from `src/content_verifier.py:276-320`.
+- u10: smoke on real sample + synth HTML; NO `src.phase_z2_pipeline` import.
+- u11: md captures U2 wiring contract — hook points, frame-contract pattern redesign (no Phase Q area reuse), IMP-07 gate. Resolves codex r1 `missing_unit` w/o Stage 1 scope_lock breach.
+
+### Out-of-scope
+- B-2 main reverse path (IMP-07).
+- Direct `src/content_verifier.py` import from any `phase_z2_*` module.
+- `REQUIRED_PATTERNS` value reuse.
+- `FORBIDDEN_KEI_MEMOS` / `generate_with_retry` (H4/H5 archive).
+- Step 14 image_aspect/table overflow (IMP-15).
+- u11 implementation (md only — no `.py`, no test, no wiring).
+- IMP-16-U2 code wiring (blocked behind IMP-07 hard-link gate).
+
+### Rollback
+Per-unit revert. u1-u10 = append-only to one new file (no cross-file edits). u11 = doc-only, zero code impact.
+
+=== EVIDENCE ===
+- Stage 1 exit report (commit 614c533): U1 = free_to_land; U2 = blocked_until_imp07_lands.
+- Codex r2 verdict: all u1-u11 = ok (each ≤50 lines AND ≤3 files).
+- Codex r1 `missing_unit: IMP-16-U2` → addressed by u11; Codex r2 ok.
+- `src/content_verifier.py` symbols verified Stage 1 (lines 29-490; all ported targets present).
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67`: IMP-16 ↓ low, hard link IMP-07.
+- Repo grep `html_to_slide_mdx|edited_html_to_mdx|reverse_path` → 0 hits (IMP-07 pending; U2 gate intact).
+- u11 path = new file, no overwrite risk.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_simulation-plan_codex_r1.md

@@ -0,0 +1,112 @@
+=== IMPLEMENTATION_UNITS ===
+reviewed_units:
+  - id: u1
+    summary: VerificationResult dataclass/module anchor.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_dataclass.py]
+    tests: [tests/phase_z2/test_pz2_vu_dataclass.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u2
+    summary: _TextExtractor + extract_text_from_html.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_extract.py]
+    tests: [tests/phase_z2/test_pz2_vu_extract.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u3
+    summary: _PARTICLES, _ENDING_NORMALIZE, normalize_for_comparison.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_normalize.py]
+    tests: [tests/phase_z2/test_pz2_vu_normalize.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u4
+    summary: extract_keywords.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_keywords.py]
+    tests: [tests/phase_z2/test_pz2_vu_keywords.py]
+    estimate_lines: 20
+    verdict: ok
+  - id: u5
+    summary: _META_PREFIXES + strip_meta_lines.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_meta_strip.py]
+    tests: [tests/phase_z2/test_pz2_vu_meta_strip.py]
+    estimate_lines: 32
+    verdict: ok
+  - id: u6
+    summary: split_into_sentences.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_sentence_split.py]
+    tests: [tests/phase_z2/test_pz2_vu_sentence_split.py]
+    estimate_lines: 28
+    verdict: ok
+  - id: u7
+    summary: _match_sentence_against_html.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_match_helper.py]
+    tests: [tests/phase_z2/test_pz2_vu_match_helper.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u8
+    summary: verify_text_preservation.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_preservation.py]
+    tests: [tests/phase_z2/test_pz2_vu_preservation.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u9
+    summary: detect_invented_text.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_invented.py]
+    tests: [tests/phase_z2/test_pz2_vu_invented.py]
+    estimate_lines: 42
+    verdict: ok
+  - id: u10
+    summary: E2E smoke against sample MDX and synthetic HTML.
+    files: [tests/phase_z2/test_pz2_vu_integration.py]
+    tests: [tests/phase_z2/test_pz2_vu_integration.py]
+    estimate_lines: 40
+    verdict: ok
+missing_unit:
+  id: IMP-16-U2
+  summary: Blocked design-only plan for Step 1/2/14/21/22 wiring after IMP-07 lands.
+  required_files: [src/phase_z2_pipeline.py, src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_reverse_verification_wiring.py]
+  required_tests: [tests/phase_z2/test_pz2_reverse_verification_wiring.py]
+  required_estimate_lines: "<=50 per unit"
+  verdict: missing
+
+Rationale:
+- u1-u10 are small enough: every listed unit is <=50 estimated lines and <=3 files.
+- The plan covers the free-to-land H3 deterministic utility port and avoids direct src/content_verifier.py import.
+- The issue body axes Step 1, Step 2, Step 14, Step 21, and Step 22 are not covered with full per-axis detail; a blocked U2 design unit is required even if implementation stays gated by IMP-07.
+
+Out-of-scope notes:
+- IMP-07 main reverse path remains out of scope and pending.
+- REQUIRED_PATTERNS values, FORBIDDEN_KEI_MEMOS, Kei/AI retry, and IMP-15 visual overflow/aspect work remain out of scope.
+- No sample-specific hardcoding should be accepted; tests may use samples but must assert general deterministic behavior.
+
+Rollback strategy:
+- Rewind Stage 2 plan only; no code edits should land from this stage. If Stage 3 later starts, revert by removing the new utility module and its specific tests only.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- rg -n "IMP-16|IMP-07|html_to_slide_mdx|edited_html_to_mdx|reverse_path" docs src tests -S
+- rg -n "^(class VerificationResult|class _TextExtractor|def extract_text_from_html|_PARTICLES|_ENDING_NORMALIZE|def normalize_for_comparison|def extract_keywords|def strip_meta_lines|def split_into_sentences|def verify_text_preservation|def detect_invented_text|REQUIRED_PATTERNS|def verify_structure|def verify_area|def verify_all_areas|FORBIDDEN_KEI_MEMOS|def verify_no_forbidden_content)" src/content_verifier.py
+- rg --files tests src docs | rg "(phase_z2|content_verifier|PHASE-Z|PHASE-Q|mdx_batch)"
+- rg -n "content_verifier" src tests docs -S
+- rg -n "Step 1|Step 2|Step 14|Step 21|Step 22|debug\.json|trace|visual_check" docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs\architecture\PHASE-Q-AUDIT.md
+- Get-ChildItem -LiteralPath 'samples\mdx_batch' -Filter '*.mdx' | Select-Object -ExpandProperty Name
+- pytest -q tests
+Files checked:
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- docs/architecture/PHASE-Q-AUDIT.md
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md
+- src/content_verifier.py
+- src/phase_z2_pipeline.py
+- tests/phase_z2/
+Test results:
+- pytest -q tests: 200 passed in 40.37s
+Verified facts:
+- Backlog keeps IMP-07 pending at docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:51.
+- Backlog keeps IMP-16 pending and hard-linked to IMP-07 at docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67.
+- The H3 deterministic source symbols are present in src/content_verifier.py at the expected symbol lines.
+- No src/phase_z2_* file currently imports src.content_verifier.
+- Existing Phase Z docs map B-2 verification helper to Step 1/2/14/21/22.
+- samples/mdx_batch contains 01.mdx through 05.mdx.
+rewind_target: stage_2_plan
+FINAL_CONSENSUS: NO

.orchestrator/drafts/16_stage_simulation-plan_codex_r2.md

@@ -0,0 +1,114 @@
+=== IMPLEMENTATION_UNITS ===
+reviewed_units:
+  - id: u1
+    summary: VerificationResult dataclass/module anchor.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_dataclass.py]
+    tests: [tests/phase_z2/test_pz2_vu_dataclass.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u2
+    summary: _TextExtractor + extract_text_from_html.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_extract.py]
+    tests: [tests/phase_z2/test_pz2_vu_extract.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u3
+    summary: _PARTICLES, _ENDING_NORMALIZE, normalize_for_comparison.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_normalize.py]
+    tests: [tests/phase_z2/test_pz2_vu_normalize.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u4
+    summary: extract_keywords.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_keywords.py]
+    tests: [tests/phase_z2/test_pz2_vu_keywords.py]
+    estimate_lines: 20
+    verdict: ok
+  - id: u5
+    summary: _META_PREFIXES + strip_meta_lines.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_meta_strip.py]
+    tests: [tests/phase_z2/test_pz2_vu_meta_strip.py]
+    estimate_lines: 32
+    verdict: ok
+  - id: u6
+    summary: split_into_sentences.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_sentence_split.py]
+    tests: [tests/phase_z2/test_pz2_vu_sentence_split.py]
+    estimate_lines: 28
+    verdict: ok
+  - id: u7
+    summary: _match_sentence_against_html.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_match_helper.py]
+    tests: [tests/phase_z2/test_pz2_vu_match_helper.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u8
+    summary: verify_text_preservation.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_preservation.py]
+    tests: [tests/phase_z2/test_pz2_vu_preservation.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u9
+    summary: detect_invented_text.
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_invented.py]
+    tests: [tests/phase_z2/test_pz2_vu_invented.py]
+    estimate_lines: 42
+    verdict: ok
+  - id: u10
+    summary: E2E smoke against sample MDX and synthetic HTML; no pipeline import.
+    files: [tests/phase_z2/test_pz2_vu_integration.py]
+    tests: [tests/phase_z2/test_pz2_vu_integration.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u11
+    summary: Design-only IMP-16-U2 wiring contract for Step 1/2/14/21/22, frame-contract pattern redesign, and IMP-07 gate; no Python.
+    files: [docs/architecture/IMP-16-U2-WIRING-DESIGN.md]
+    tests: []
+    estimate_lines: 50
+    verdict: ok
+
+Rationale:
+- u1: establishes the new Phase Z-owned utility surface; no src.content_verifier import.
+- u2: text extraction uses stdlib HTMLParser and remains pure/deterministic.
+- u3: Korean normalization constants are text-axis utilities, not Phase Q frame patterns.
+- u4: keyword extraction composes normalization and keeps the port bounded.
+- u5: meta-line stripping is analyzer-text cleanup, not REQUIRED_PATTERNS reuse.
+- u6: sentence split composes u5 and stays content-agnostic.
+- u7: private matcher isolates ratio/fuzzy logic so u8 remains small.
+- u8: preservation verifier composes prior utilities and preserves the deterministic threshold behavior.
+- u9: invented-text detector remains pure regex/string logic; no AI or Kei path.
+- u10: smoke coverage is additive and must assert general behavior, not sample-specific output.
+- u11: covers the issue-body Step 1/2/14/21/22 axes as blocked design, satisfying IMP-07 hard-link without wiring.
+
+Out-of-scope notes:
+- IMP-07 main reverse path remains out of scope and pending.
+- No Step 1/2/14/21/22 runtime wiring lands before IMP-07.
+- REQUIRED_PATTERNS values, verify_structure/verify_area/verify_all_areas, FORBIDDEN_KEI_MEMOS, verify_no_forbidden_content, generate_with_retry, Selenium/Vision, and IMP-15 visual overflow/aspect work stay out of scope.
+
+Rollback strategy:
+- Stage 2 has no code edits. Later Stage 3 rollback is per-unit revert: remove the new utility module/tests for u1-u10 or the doc-only u11 file.
+
+=== EVIDENCE ===
+Commands run:
+- Get-Content -Raw -LiteralPath D:\ad-hoc\kei\design_agent\.orchestrator\tmp\codex_prompt.txt
+- git status --short
+- rg -n "IMP-16|IMP-07|html_to_slide_mdx|edited_html_to_mdx|reverse_path|phase_z2_verification_utils|content_verifier" docs src tests samples -S
+- rg -n "^class VerificationResult|^class _TextExtractor|^def extract_text_from_html|^_PARTICLES|^_ENDING_NORMALIZE|^def normalize_for_comparison|^def extract_keywords|^def strip_meta_lines|^def split_into_sentences|^def verify_text_preservation|^def detect_invented_text|^FORBIDDEN_KEI_MEMOS|^def verify_no_forbidden_content|^REQUIRED_PATTERNS|^def verify_structure|^def verify_area|^def verify_all_areas" src/content_verifier.py
+- Test-Path src/phase_z2_verification_utils.py; Test-Path docs/architecture/IMP-16-U2-WIRING-DESIGN.md; Get-ChildItem tests/phase_z2 -Filter test_pz2_vu_*.py
+- pytest -q tests
+Files checked:
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- docs/architecture/PHASE-Q-AUDIT.md
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- src/content_verifier.py
+- tests/phase_z2/
+Test results:
+- pytest -q tests: 200 passed in 40.29s
+Verified facts:
+- IMP-07 is still pending at docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:51.
+- IMP-16 is pending and hard-linked to IMP-07 at docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67.
+- H3 deterministic source symbols exist in src/content_verifier.py at the expected symbol lines 29-490.
+- Phase Z verification utility files do not exist yet, so the r2 plan is additive.
+- u1-u11 each list <=3 files and <=50 estimated lines.
+- u11 now covers the previously missing blocked IMP-16-U2 Step 1/2/14/21/22 design axis without pipeline wiring.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_simulation-plan_codex_r3.md

@@ -0,0 +1,123 @@
+[Codex #3] Stage 2 r3 verification for IMP-16. I verified the r3 plan against the issue body axes, Stage 1 binding contract, local source anchors, unit sizing rule, and required baseline tests. The plan is complete for Stage 2.
+
+=== IMPLEMENTATION_UNITS ===
+reviewed_units:
+  - id: u1
+    covers: IMP-16-U1 module anchor and VerificationResult
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_dataclass.py]
+    tests: [tests/phase_z2/test_pz2_vu_dataclass.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u2
+    covers: pure HTML text extraction
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_extract.py]
+    tests: [tests/phase_z2/test_pz2_vu_extract.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u3
+    covers: Korean text normalization constants and function
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_normalize.py]
+    tests: [tests/phase_z2/test_pz2_vu_normalize.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u4
+    covers: keyword extraction
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_keywords.py]
+    tests: [tests/phase_z2/test_pz2_vu_keywords.py]
+    estimate_lines: 20
+    verdict: ok
+  - id: u5
+    covers: meta-line stripping
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_meta_strip.py]
+    tests: [tests/phase_z2/test_pz2_vu_meta_strip.py]
+    estimate_lines: 32
+    verdict: ok
+  - id: u6
+    covers: sentence splitting
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_sentence_split.py]
+    tests: [tests/phase_z2/test_pz2_vu_sentence_split.py]
+    estimate_lines: 28
+    verdict: ok
+  - id: u7
+    covers: private sentence-vs-html match helper
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_match_helper.py]
+    tests: [tests/phase_z2/test_pz2_vu_match_helper.py]
+    estimate_lines: 30
+    verdict: ok
+  - id: u8
+    covers: verify_text_preservation
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_preservation.py]
+    tests: [tests/phase_z2/test_pz2_vu_preservation.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u9
+    covers: detect_invented_text
+    files: [src/phase_z2_verification_utils.py, tests/phase_z2/test_pz2_vu_invented.py]
+    tests: [tests/phase_z2/test_pz2_vu_invented.py]
+    estimate_lines: 42
+    verdict: ok
+  - id: u10
+    covers: sample-backed smoke without pipeline import
+    files: [tests/phase_z2/test_pz2_vu_integration.py]
+    tests: [tests/phase_z2/test_pz2_vu_integration.py]
+    estimate_lines: 40
+    verdict: ok
+  - id: u11
+    covers: IMP-16-U2 design-only contract for Step 1/2/14/21/22, redesigned frame-contract patterns, trace/report hooks, and IMP-07 gate
+    files: [docs/architecture/IMP-16-U2-WIRING-DESIGN.md]
+    tests: []
+    estimate_lines: 50
+    verdict: ok
+
+Rationale:
+- u1: establishes a Phase Z-owned deterministic surface; no src.content_verifier import.
+- u2: text extraction is stdlib-only and covers the B-2 HTML side.
+- u3: normalization is text-axis utility scope, not Phase Q REQUIRED_PATTERNS reuse.
+- u4: keyword extraction is needed by preservation and invented-text checks.
+- u5: analyzer meta stripping is orthogonal to frame-contract validation.
+- u6: sentence splitting composes the meta-strip utility for MDX text checks.
+- u7: helper split keeps the preservation unit small and testable.
+- u8: text preservation covers Step 2 verification semantics for re-entered MDX.
+- u9: invented text detection covers the reverse-path hallucination guard.
+- u10: smoke coverage uses existing sample MDX but does not hardcode sample passing as the pipeline rule.
+- u11: covers the blocked IMP-16-U2 axis without merging Step 1/2/14/21/22 wiring before IMP-07.
+
+Out-of-scope notes:
+- IMP-07 main reverse path remains out of scope and pending.
+- Step 14 image/table overflow enhancement remains IMP-15 territory.
+- Kei/AI retry loops and FORBIDDEN_KEI_MEMOS remain archive-only and are not ported.
+- REQUIRED_PATTERNS values are not reused; only the deterministic checking approach is referenced.
+
+Rollback strategy:
+- Revert only src/phase_z2_verification_utils.py, the listed tests, and docs/architecture/IMP-16-U2-WIRING-DESIGN.md if Stage 3 implementation regresses. U2 has no runtime rollback because it is design-only.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- rg -n "IMP-16|IMP-07|content_verifier|phase_z2_verification_utils|html_to_slide_mdx|edited_html_to_mdx|reverse_path" docs src tests samples -g "*.md" -g "*.py"
+- rg --files docs src tests samples | rg "(PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG|PHASE-Q-AUDIT|PHASE-Q-INSIGHT-TO-22STEP-MAP|PHASE-Z-PIPELINE-OVERVIEW|content_verifier|phase_z2|mdx_batch)"
+- rg -n "^(class |def |REQUIRED_PATTERNS|FORBIDDEN_KEI_MEMOS|_PARTICLES|_ENDING_NORMALIZE|_META_PREFIXES)" src/content_verifier.py
+- rg -n "Step 1|Step 2|Step 14|Step 21|Step 22|debug|visual_check|trace|report" docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md src/phase_z2_pipeline.py
+- pytest -q tests
+
+Files checked:
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- docs/architecture/PHASE-Q-AUDIT.md
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md
+- src/content_verifier.py
+- src/phase_z2_pipeline.py
+- tests/phase_z2/
+- samples/mdx_batch/03.mdx
+
+Test results:
+- pytest -q tests: 200 passed in 40.26s
+
+Verified facts:
+- Backlog keeps IMP-16 pending, low priority, hard-linked to IMP-07, and scoped to Step 1/2/14/21/22 verification support.
+- Backlog keeps IMP-07 pending; local grep found no phase_z2 reverse-path implementation for html_to_slide_mdx, edited_html_to_mdx, or reverse_path.
+- content_verifier.py contains the planned deterministic H3 symbols and also contains FORBIDDEN_KEI_MEMOS, REQUIRED_PATTERNS, and structure-area functions that remain excluded or redesigned.
+- Phase Z pipeline has existing Step 1, Step 2, Step 14, Step 21, and Step 22 anchors for future U2 design, but U2 runtime wiring is correctly gated.
+- Every implementation unit is <=50 estimated lines and touches <=3 files.
+- The plan covers text, visual/structure design, and trace/report axes while preserving AI=0 normal-path guardrails.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_test-verify_claude_r1.md

@@ -0,0 +1,120 @@
+[Claude #5] Stage 4 r1 — IMP-16-U1 test/verification round 1
+
+scope: verify Stage 3 deliverables (u1~u10 runtime code + u11 design-only md) against Stage 2 plan; targeted + regression test execution; hardcoding / forbidden-import / IMP-07 gate audit. No commit performed (Stage 3 artifacts remain untracked, per Stage 3 exit report).
+
+=== DIFF vs PLAN ===
+
+implementation_units delivered (Stage 2 plan → Stage 3 artifacts):
+
+| unit | files | LOC | status |
+|---|---|---|---|
+| u1 (anchor + VerificationResult) | src/phase_z2_verification_utils.py (lines 19-33) + tests/phase_z2/test_pz2_vu_dataclass.py | module 335 / test 59 | present |
+| u2 (extract_text_from_html) | src/phase_z2_verification_utils.py (36-73) + test_pz2_vu_extract.py | test 54 | present |
+| u3 (normalize_for_comparison + _PARTICLES + _ENDING_NORMALIZE) | src/phase_z2_verification_utils.py (76-104) + test_pz2_vu_normalize.py | test 64 | present |
+| u4 (extract_keywords) | src/phase_z2_verification_utils.py (107-126) + test_pz2_vu_keywords.py | test 52 | present |
+| u5 (strip_meta_lines) | src/phase_z2_verification_utils.py (129-166) + test_pz2_vu_meta_strip.py | test 73 | present |
+| u6 (split_into_sentences) | src/phase_z2_verification_utils.py (169-198) + test_pz2_vu_sentence_split.py | test 69 | present |
+| u7 (_sentence_matches_html) | src/phase_z2_verification_utils.py (201-235) + test_pz2_vu_match_helper.py | test 66 | present |
+| u8 (verify_text_preservation) | src/phase_z2_verification_utils.py (238-285) + test_pz2_vu_preservation.py | test 119 | present |
+| u9 (detect_invented_text) | src/phase_z2_verification_utils.py (288-335) + test_pz2_vu_invented.py | test 84 | present |
+| u10 (integration smoke) | tests/phase_z2/test_pz2_vu_integration.py | 106 | present |
+| u11 (design-only contract) | docs/architecture/IMP-16-U2-WIRING-DESIGN.md | 75 | present |
+
+artifact status: all 12 files untracked (git status, scoped to IMP-16 paths) — Stage 3 exit-report contract preserved. No modifications to tracked source files in IMP-16 scope. The only `git diff --stat` hits (orchestrator.py, tests/orchestrator_unit/test_subprocess_cleanup.py) are out of IMP-16 scope (carried over from commit `f3bff89 feat(orchestrator): initial orchestrator + subprocess cleanup hardening` baseline, pre-IMP-16).
+
+=== TARGETED TESTS (u1~u10) ===
+
+```
+python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py \
+  tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py \
+  tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py \
+  tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py \
+  tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_integration.py -v
+```
+
+result: 56 passed in 0.19s.
+
+per-unit pass count (matches Stage 2 unit decomposition):
+- u1 dataclass: 4 passed (incl. AST guard `test_module_importable_without_content_verifier`)
+- u2 extract: 5 passed
+- u3 normalize: 6 passed
+- u4 keywords: 6 passed
+- u5 meta_strip: 5 passed
+- u6 sentence_split: 6 passed
+- u7 match_helper: 5 passed
+- u8 preservation: 7 passed (incl. threshold lock + truncation + cap-at-limit)
+- u9 invented: 7 passed (incl. constants lock + 80-char truncation)
+- u10 integration: 5 passed (incl. AST-based forbidden-import scan)
+
+=== REGRESSION ===
+
+```
+python -m pytest tests/phase_z2/ tests/orchestrator_unit/test_subprocess_cleanup.py -q
+```
+
+result: 172 passed in 33.69s. delta vs Stage 2 baseline = +56 (the new u1~u10 tests). No prior `tests/phase_z2/*` test regressed; the recently-modified subprocess-cleanup guard suite (commit `f3bff89`) remains green.
+
+=== HARDCODING AUDIT (RULE 0 + RULE 7) ===
+
+- `grep -nE "MDX 03|MDX 04|MDX 05|03\\.mdx|04\\.mdx|05\\.mdx"` against src/phase_z2_verification_utils.py + tests/phase_z2/test_pz2_vu_*.py → 0 hits.
+- u10 integration uses `samples/mdx_batch/02.mdx` only as a smoke fixture; the file's docstring + Stage 2 u10 rationale explicitly note "smoke coverage uses existing sample MDX but does not hardcode sample passing as the pipeline rule." Asserts are API-shape only (VerificationResult instance, area_name passthrough, `0.0 ≤ score ≤ 1.0`, list return) for the real-sample case; the mirrored-HTML pass test uses synthetic round-trip input, not a hardcoded score floor against MDX 02.
+- All numeric thresholds (`_TEXT_PRESERVATION_DEFAULT_THRESHOLD = 0.70`, `_SENTENCE_KEYWORD_MATCH_THRESHOLD = 0.6`, `_SENTENCE_SEQUENCE_MATCH_THRESHOLD = 0.65`, `_INVENTED_TEXT_KEYWORD_THRESHOLD = 0.4`, `_INVENTED_TEXT_MIN_LENGTH = 15`, `_INVENTED_TEXT_TRUNCATE_LEN = 80`, `_MISSING_SENTENCE_REPORT_LIMIT = 5`, `_MISSING_SENTENCE_TRUNCATE_LEN = 60`) lifted to named module constants and locked by u7/u8/u9 tests.
+- No sample-specific behavior baked into module logic; tokens/patterns are Korean-locale, not sample-locale.
+
+=== FORBIDDEN-IMPORT / SCOPE-LOCK AUDIT (Stage 1 + Stage 2 binding contracts) ===
+
+| forbidden surface | check | result |
+|---|---|---|
+| `src.content_verifier` import (any form) | `grep -n "content_verifier" src/phase_z2_verification_utils.py tests/phase_z2/test_pz2_vu_*.py` → only docstrings + AST-guard test assertions. AST scan in u1 + u10 actively rejects this import. | not imported |
+| `FORBIDDEN_KEI_MEMOS` / `verify_no_forbidden_content` (H4) | `grep -n "FORBIDDEN_KEI"` → 0 hits in U1 source + tests | not ported |
+| `generate_with_retry` (H5 archive) | `grep -n "generate_with_retry"` → 0 hits in U1 source + tests | not ported |
+| `REQUIRED_PATTERNS` values reuse | `grep -n "REQUIRED_PATTERNS"` → only one comment line in test_pz2_vu_normalize.py noting Phase Q surface size; no value copy | not reused |
+| `verify_structure` / `verify_area` / `verify_all_areas` | `grep -n "verify_structure\|verify_area\|verify_all_areas"` → 0 hits in U1 source + tests | not ported |
+| AI / Kei / httpx / SSE coupling | `grep -nE "anthropic\|kei_client\|httpx\|requests\|sse"` → 0 hits | none |
+| `src/` runtime wiring of U1 module | `grep -rn "phase_z2_verification_utils" src/` → 0 hits (no other src/ module imports U1) | not wired |
+
+=== IMP-07 HARD GATE PRESERVED ===
+
+- `grep -rn "html_to_slide_mdx\|edited_html_to_mdx\|reverse_path\|reverse-path" src/ --include="*.py"` → 1 hit, and that hit is the U1 module docstring text ("compare reverse-path HTML against MDX text"), NOT a runtime symbol or callsite. IMP-07 main reverse path remains absent in runtime code.
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md` §"Gate (hard block — do not merge wiring before this clears)" requires IMP-07 = `implemented` + `verified` and at least one runtime hit of those greps under `src/` before any of the Step 1/2/14/21/22 wiring described in §"Per-step wiring contract" lands. Both conditions remain unsatisfied; U2 wiring correctly stays design-only.
+- The four reserved debug.json keys (`pipeline.reverse_path_reentry`, `step01.reentry_sentence_count`, `step02.reentry_text_preservation`, `step14.reentry_invented_text_fragments`) are reserved in u11 but NOT yet present in any runtime schema (verified by absence of corresponding write sites under `src/`).
+
+=== GUARDRAIL RESIDUAL ===
+
+- AI-isolation contract (`feedback_ai_isolation_contract`): respected — module is pure-deterministic, stdlib-only (`re`, `dataclasses`, `difflib`, `html.parser`).
+- Auto-pipeline-first (`feedback_auto_pipeline_first`): u11 design explicitly forbids `review_required` / `review_queue` injection; deterministic adapter-abort on preservation fail.
+- Spacing direction (`feedback_phase_z_spacing_direction`): N/A — verification axis, no layout / CSS / padding shrink path.
+- No-hardcoding (`feedback_no_hardcoding`): see Hardcoding Audit above.
+- 1-turn = 1-step (`feedback_one_step_per_turn`): Stage 4 r1 is a single verification round; no Stage 5 recommendation made.
+- Scope-qualified verification (`feedback_scope_qualified_verification`): every pass count above carries unit / test-file / command scope.
+- Status semantics (`feedback_artifact_status_naming`): u11 explicitly states preservation/invented-text are *telemetry* and do not flip Step 20 `compute_slide_status`.
+
+=== VERDICT ===
+
+PASS for Stage 4 round 1.
+
+- Stage 3 deliverables match Stage 2 plan unit-for-unit (u1~u11).
+- 56/56 targeted tests pass; 172/172 broader Phase Z + subprocess-cleanup regression suite pass.
+- No hardcoding, no forbidden Phase Q imports, no AI/Kei coupling, no IMP-07 gate violation, no schema regression (U2 wiring not yet present in runtime).
+- Artifacts remain untracked per Stage 3 exit report; staging / commit decision belongs to Stage 5.
+
+=== EVIDENCE ===
+
+commands run:
+- `git status --short` — confirmed 12 IMP-16 paths (1 module + 1 design md + 10 test files) untracked; modified tracked files (`orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py`) carry over from baseline commit `f3bff89` and are out of IMP-16 scope.
+- `wc -l src/phase_z2_verification_utils.py docs/architecture/IMP-16-U2-WIRING-DESIGN.md tests/phase_z2/test_pz2_vu_*.py` — 335 / 75 / 706 (10 test files) = 1116 LOC across artifacts.
+- `python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_integration.py -v` — 56 passed in 0.19s.
+- `python -m pytest tests/phase_z2/ tests/orchestrator_unit/test_subprocess_cleanup.py -q` — 172 passed in 33.69s.
+- `python -c "from src.phase_z2_verification_utils import VerificationResult, extract_text_from_html, normalize_for_comparison, extract_keywords, strip_meta_lines, split_into_sentences, verify_text_preservation, detect_invented_text; print('OK')"` — OK (full public surface importable in isolation).
+- `grep -rn "phase_z2_verification_utils" src/ --include="*.py"` — 0 hits (no runtime wiring leak under src/).
+- `grep -rn "html_to_slide_mdx\|edited_html_to_mdx\|reverse_path\|reverse-path" src/ --include="*.py"` — 1 hit, all in the U1 module docstring (not a runtime symbol). IMP-07 main reverse path still absent.
+- `grep -nE "content_verifier|FORBIDDEN_KEI|generate_with_retry|httpx|sse|REQUIRED_PATTERNS|verify_structure|verify_area"` against U1 module + tests — only docstring/comment/test-guard references; zero imports, zero symbol reuse.
+
+commit_sha referenced for baseline: `614c53358ecaa92e9f6fc979b2399da028e15d41` (HEAD; pre-Stage-3 artifacts not yet committed).
+
+source_comment_ids:
+- [Claude #4] Stage 2 r3 plan (u1~u11)
+- [Codex #26] Stage 3 r11 verification of u11
+- [Claude #5] Stage 4 r1 verification (this comment)
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/16_stage_test-verify_codex_r1.md

@@ -0,0 +1,61 @@
+[Codex #27] Stage 4 r1 independent verification for IMP-16
+
+Scope: verified Stage 3 deliverables for IMP-16-U1 (u1-u10 runtime utility + tests) and u11 design-only wiring contract. No commit performed. The IMP-16 files remain untracked and ready for a commit gate decision.
+
+Result: PASS. Commit OK for the scoped IMP-16 artifacts only.
+
+Diff / scope verification:
+- Present IMP-16 artifacts: `src/phase_z2_verification_utils.py`, `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`, and 10 `tests/phase_z2/test_pz2_vu_*.py` files.
+- `git status --short -- <IMP-16 paths>` shows all 12 scoped files as untracked.
+- `git diff --stat` shows tracked modifications only in `orchestrator.py` and `tests/orchestrator_unit/test_subprocess_cleanup.py`; those are outside the IMP-16 commit scope.
+- Runtime grep under `src/` for `html_to_slide_mdx|edited_html_to_mdx|reverse_path|reverse-path|reentry_|re-entry` found only one docstring mention in `src/phase_z2_verification_utils.py`; no Step 1/2/14/21/22 wiring landed.
+
+Guardrail verification:
+- `src/phase_z2_verification_utils.py` imports only stdlib modules: `re`, `dataclasses`, `difflib`, and `html.parser`.
+- Corrected import audit across the module and tests found no `src.content_verifier` import and no Phase Q runtime dependency.
+- Forbidden-coupling grep found no runtime use of `FORBIDDEN_KEI_MEMOS`, `generate_with_retry`, `httpx`, `SSE`, `REQUIRED_PATTERNS`, `verify_structure`, `verify_area`, or `verify_all_areas`; only design/reference text appears in the u11 markdown and test comments.
+- `python -m py_compile src/phase_z2_verification_utils.py` passed. The generated bytecode artifact was removed afterward.
+
+Test results:
+- IMP-16 targeted suite: 56 passed in 0.19s.
+- Regression guard from the Stage 2 baseline: 15 passed in 7.59s.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `rg --files src tests docs | rg "(phase_z2_verification_utils|test_pz2_vu_|IMP-16-U2-WIRING-DESIGN)"`
+- `rg -n "src\.content_verifier|content_verifier|FORBIDDEN_KEI_MEMOS|generate_with_retry|httpx|SSE|REQUIRED_PATTERNS|verify_structure|verify_area|verify_all_areas|html_to_slide_mdx|edited_html_to_mdx|reverse_path" src/phase_z2_verification_utils.py tests/phase_z2 docs/architecture/IMP-16-U2-WIRING-DESIGN.md`
+- `git diff --stat`
+- `python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_integration.py -v`
+- `python -m pytest tests/orchestrator_unit/test_subprocess_cleanup.py -v`
+- `git status --short -- src/phase_z2_verification_utils.py docs/architecture/IMP-16-U2-WIRING-DESIGN.md tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_integration.py`
+- `rg -n "html_to_slide_mdx|edited_html_to_mdx|reverse_path|reverse-path|reentry_|re-entry" src --glob "*.py"`
+- `rg -n "^import |^from " src/phase_z2_verification_utils.py tests/phase_z2 --glob "test_pz2_vu_*.py"`
+- `python -m py_compile src/phase_z2_verification_utils.py`
+
+Files checked:
+- `src/phase_z2_verification_utils.py`
+- `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`
+- `tests/phase_z2/test_pz2_vu_dataclass.py`
+- `tests/phase_z2/test_pz2_vu_extract.py`
+- `tests/phase_z2/test_pz2_vu_normalize.py`
+- `tests/phase_z2/test_pz2_vu_keywords.py`
+- `tests/phase_z2/test_pz2_vu_meta_strip.py`
+- `tests/phase_z2/test_pz2_vu_sentence_split.py`
+- `tests/phase_z2/test_pz2_vu_match_helper.py`
+- `tests/phase_z2/test_pz2_vu_preservation.py`
+- `tests/phase_z2/test_pz2_vu_invented.py`
+- `tests/phase_z2/test_pz2_vu_integration.py`
+- `tests/orchestrator_unit/test_subprocess_cleanup.py`
+
+Test results:
+- `python -m pytest tests/phase_z2/test_pz2_vu_dataclass.py tests/phase_z2/test_pz2_vu_extract.py tests/phase_z2/test_pz2_vu_normalize.py tests/phase_z2/test_pz2_vu_keywords.py tests/phase_z2/test_pz2_vu_meta_strip.py tests/phase_z2/test_pz2_vu_sentence_split.py tests/phase_z2/test_pz2_vu_match_helper.py tests/phase_z2/test_pz2_vu_preservation.py tests/phase_z2/test_pz2_vu_invented.py tests/phase_z2/test_pz2_vu_integration.py -v` -> 56 passed in 0.19s.
+- `python -m pytest tests/orchestrator_unit/test_subprocess_cleanup.py -v` -> 15 passed in 7.59s.
+
+Verified facts:
+- IMP-16-U1 utility surface is present in `src/phase_z2_verification_utils.py` and remains deterministic/stdlib-only.
+- IMP-16-U2 remains design-only in `docs/architecture/IMP-16-U2-WIRING-DESIGN.md`.
+- No Phase Z runtime import of `src.content_verifier` was introduced by the scoped artifacts.
+- No IMP-07 reverse-path runtime wiring was introduced under `src/`.
+- Scoped IMP-16 files are untracked; no commit was made in Stage 4.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_code-edit_claude_r1.md

@@ -0,0 +1,36 @@
+Executing unit: u1
+
+[Claude #4] Stage 3 Round #1 — unit u1 (comment anchor correction).
+
+**unit_executed**: u1
+
+**files_changed**:
+- `src/phase_z2_pipeline.py` (1 line modified, line 564)
+- `tests/orchestrator_unit/test_imp17_comment_anchor.py` (new file, 27 lines)
+
+**diff_summary**:
+- `src/phase_z2_pipeline.py:564` — comment `restructure  → AI-assisted frame-aware adaptation (deferred to IMP-31)` → `restructure  → AI-assisted frame-aware adaptation (deferred to IMP-17 — carve-out, AI fallback only, normal path 밖)`. Verbatim IMP-17 (Stage 2 R2 lock Q1).
+- Line 565 IMP-29 reject route comment untouched per scope_lock.
+- New anchor test `tests/orchestrator_unit/test_imp17_comment_anchor.py` with two cases:
+  1. `test_line_564_references_imp17_not_imp31` — asserts line 564 contains `restructure` + `IMP-17` and does not contain `IMP-31`.
+  2. `test_line_565_still_references_imp29` — asserts line 565 contains `reject` + `IMP-29` (regression guard for the untouched neighbour).
+- `_IMP05_ROUTE_HINTS` dict body (`ai_adaptation_required` etc.) unchanged. Runtime behavior unchanged.
+
+**test_evidence**:
+```
+$ python -m pytest -q tests/orchestrator_unit/test_imp17_comment_anchor.py
+..                                                                       [100%]
+2 passed in 0.02s
+```
+
+**caps_respected**:
+- 3-file cap: 2 files touched (1 source + 1 new test) ✓
+- 50-line cap: 1 line in source + 27 lines in new test = 28 lines total ✓
+- AI isolation: zero runtime AI call introduced ✓
+- No new IMP ID allocated; uses existing Gitea #17 / IMP-17 slot ✓
+
+**remaining_units**: [u2, u3, u4]
+
+**follow_up_issue_candidates**: (none — u1 strictly within Stage 2 scope_lock; no out-of-scope axis surfaced).
+
+Stage 3 continues in next round with u2 (carve-out doc). Stopping per 1 turn = 1 unit rule.

.orchestrator/drafts/17_stage_code-edit_claude_r2.md

@@ -0,0 +1,39 @@
+Executing unit: u2
+
+[Claude #6] Stage 3 Round #2 — unit u2 (carve-out design doc).
+
+**unit_executed**: u2
+
+**files_changed**:
+- `docs/architecture/IMP-17-CARVE-OUT.md` (new file, 44 lines)
+
+**diff_summary**:
+- New design-only doc at `docs/architecture/IMP-17-CARVE-OUT.md` (44 lines, planned ≤48). Covers all Stage 2 R2 requirements for u2:
+  - **Status header**: carve-out, design-only, normal-path AI calls = 0, no runtime fallback lands until activation gate clears.
+  - **Source anchors** section: backlog row line 68 + INSIGHT-MAP §3 + 22-step pipeline Step 12 (lines 280-287), Step 16 (lines 318-325), Step 17 (lines 326-333) + link-only pattern refs (`src/content_editor.py:21,318` + `src/sse_utils.py:16-50`) + route hint emission site (`src/phase_z2_pipeline.py:564`).
+  - **Carve-out boundary**: Allowed = Step 12 `restructure` AI proposal at content_object → Internal Region / Frame Slot granularity AND Step 16/17 retry-exhaustion AI proposal at same granularity. Forbidden = normal-path AI calls, MDX 원문 압축 / 재작성, HTML·CSS·frame contract·layout·zone topology 직접 생성, 공통 padding/spacing/tolerance 축소, 신규 IMP ID 발급.
+  - **Activation gate (3-condition AND)**: (1) explicit user GO, (2) B4 frame_selection evidence integration complete, (3) IMP-04 (catalog 확장) + IMP-05 (V4 fallback) live. 세 조건 중 하나라도 미충족이면 design-only 상태 유지.
+  - **Pattern shape reference**: link-only into `content_editor.py:21,318` + `sse_utils.py:16-50`. Phase Q Archive Candidate ([`PHASE-Q-AUDIT.md`](docs/architecture/PHASE-Q-AUDIT.md):660-673) — port 금지. `EDITOR_PROMPT` / Kei-API endpoint = 영구 단절.
+  - **AI 격리 + Kei persona 단절 contract**: AI 출력 단위 = content_object / Internal Region / Frame Slot / restructuring proposal. HTML 구조 / 레이아웃 / 프리셋 결정 금지. Phase Q persona 자산 단절, 활성 시 별도 prompt / endpoint 설계.
+- No runtime code change. No test added (u2 is docs-only per Stage 2 plan).
+- Stage 1 unresolved questions Q1/Q2/Q3 all resolved verbatim in u2 (verbatim IMP-17 ✓, 3-condition AND gate ✓, link-only pattern refs ✓).
+
+**caps_respected**:
+- 3-file cap: 1 file touched ✓
+- 50-line cap: 44 lines (≤ planned 48, ≤ cap 50) ✓
+- AI isolation contract: zero runtime AI call introduced ✓
+- No new IMP ID allocated; uses IMP-17 slot verbatim ✓
+- Kei persona endpoint remains disconnected (pattern shape link-only) ✓
+- Phase Z spacing direction guardrail explicitly cited in forbidden list ✓
+
+**verification (manual, docs-only)**:
+- `wc -l docs/architecture/IMP-17-CARVE-OUT.md` → 44 lines.
+- `grep -n "IMP-17|IMP-04|IMP-05|IMP-29"` confirms cross-refs in place, no spurious IMP-31 / IMP-29 reject confusion.
+- All linked source paths (`src/phase_z2_pipeline.py:564`, `src/content_editor.py:21,318`, `src/sse_utils.py:16-50`) verified present in repo at quoted line numbers during this round.
+- All linked architecture docs (`PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, `PHASE-Q-INSIGHT-TO-22STEP-MAP.md`, `PHASE-Z-PIPELINE-OVERVIEW.md`, `PHASE-Q-AUDIT.md`) exist in same directory (`docs/architecture/`).
+
+**remaining_units**: [u3, u4]
+
+**follow_up_issue_candidates**: (none — u2 strictly within Stage 2 scope_lock; no out-of-scope axis surfaced).
+
+Stage 3 continues in next round with u3 (backlog row carve-out link + gate pointer). Stopping per 1 turn = 1 unit rule. No commit, no push.

.orchestrator/drafts/17_stage_code-edit_claude_r3.md

@@ -0,0 +1,39 @@
+Executing unit: u3
+
+[Claude #8] Stage 3 Round #3 — unit u3 (backlog row carve-out link + activation gate pointer).
+
+**unit_executed**: u3
+
+**files_changed**:
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (1 row modified, line 68 — IMP-17 row scope cell)
+
+**diff_summary**:
+- Backlog IMP-17 row (line 68) scope cell prepended with:
+  - Markdown link `[carve-out boundary + activation gate](IMP-17-CARVE-OUT.md)` — relative link resolves within `docs/architecture/` directory.
+  - Short 3-cond AND pointer: `User GO ∧ B4 frame_selection evidence ∧ IMP-04/05 live` with `full def in u2 doc` disclaimer routing readers to the binding definition.
+- Existing scope text retained verbatim after the new link/pointer (`` `httpx` + SSE streaming + retry + JSON parse pattern reference — light_edit / restructure proposal ``).
+- ID column, related step, source, priority, guardrail, dependency, status columns UNCHANGED.
+- Adjacent rows (IMP-16 line 67, IMP-18 line 69) UNCHANGED.
+- One logical line modified in the markdown table; net diff well under the Stage 2 4-line cap.
+- u2 carve-out doc presence verified at `docs/architecture/IMP-17-CARVE-OUT.md` (link target resolves in the same directory, so the relative path is valid).
+
+**caps_respected**:
+- 3-file cap: 1 file touched ✓
+- 50-line cap: 1 logical row modified (Stage 2 estimate = 4) ✓
+- AI isolation contract: docs-only edit, zero runtime AI call introduced ✓
+- No new IMP ID allocated; uses existing IMP-17 slot verbatim ✓
+- Activation gate full definition kept in u2 (this row carries only a pointer) — no duplication of binding text ✓
+- No-hardcoding: process-level cross-ref edit, not artifact patch ✓
+
+**verification (manual, docs-only)**:
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` confirms exactly one line changed in the backlog file (IMP-17 row).
+- `test -f docs/architecture/IMP-17-CARVE-OUT.md` returns true — the link target exists in the same directory, so the relative markdown link resolves.
+- No regression to IMP-16 (line 67), IMP-18 (line 69), or any other row.
+- The guardrail column for IMP-17 retains the binding `normal path AI 호출 0` + `Kei persona 단절` statement.
+- The dependency column for IMP-17 retains `soft link: IMP-04 + IMP-05` (matches the u2 activation gate ingredients).
+
+**remaining_units**: [u4]
+
+**follow_up_issue_candidates**: (none — u3 strictly within Stage 2 scope_lock; no out-of-scope axis surfaced).
+
+Stage 3 continues in next round with u4 (22-step map registry row prefix). Stopping per 1 turn = 1 unit rule.

.orchestrator/drafts/17_stage_code-edit_claude_r4.md

@@ -0,0 +1,37 @@
+Executing unit: u4
+
+[Claude #10] Stage 3 Round #4 — unit u4 (22-step map registry row prefix).
+
+**unit_executed**: u4
+
+**files_changed**:
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (1 logical row modified, line 123 — §3 axis registry row for AI repair fallback infra)
+
+**diff_summary**:
+- Line 123 axis label changed from `AI repair fallback infra` → `IMP-17 AI repair fallback infra (carve-out — see [`IMP-17-CARVE-OUT.md`](IMP-17-CARVE-OUT.md))`.
+- Verbatim `IMP-17` prefix (Stage 1 Q1 lock, consistent with u1 + u2 + u3).
+- Relative markdown link resolves within `docs/architecture/` (sibling file, u2 doc verified present).
+- Step column `Step 12, 16, 17`, source column `§2.6 G3 (httpx + SSE streaming + retry + JSON parse pattern)`, status column `pending`, and normal-path column `no (AI fallback only)` retained verbatim.
+- Adjacent rows (`B-2 verification 보조` line 122, `I3 SVG 좌표 보강` line 124) UNCHANGED.
+- Row 88 (Step 12 row cross-referencing `AI repair fallback infra axis → §3 registry`) intentionally LEFT UNCHANGED — Stage 2 plan scope = registry row only; the §3 reference still resolves because the registry row continues to contain the original `AI repair fallback infra` phrase.
+- `git diff --numstat` confirms `1 1 docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (1 line removed, 1 line added). Net diff well under Stage 2 3-line cap.
+
+**caps_respected**:
+- 3-file cap: 1 file touched ✓
+- 50-line cap: 1 logical row modified (Stage 2 estimate = 3) ✓
+- AI isolation contract: docs-only edit, zero runtime AI call introduced ✓
+- No new IMP ID allocated; uses existing IMP-17 slot verbatim ✓
+- normal-path column unchanged (`no (AI fallback only)`) — carve-out contract preserved ✓
+- No-hardcoding: process-level cross-ref edit, not artifact patch ✓
+
+**verification (manual, docs-only)**:
+- `git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` confirms exactly one logical row changed.
+- `ls docs/architecture/IMP-17-CARVE-OUT.md` returns the file (u2 deliverable) — the relative markdown link resolves.
+- `git diff --numstat` = `1 1` (delete + add of the same row) — no spurious whitespace or trailing edits.
+- §0 (Kei artifact disconnect) untouched. §2.6 G3 source ref untouched. §3 registry header untouched.
+
+**remaining_units**: [] (all Stage 2 implementation_units now executed: u1 ✓, u2 ✓, u3 ✓, u4 ✓).
+
+**follow_up_issue_candidates**: (none — u4 strictly within Stage 2 scope_lock; no out-of-scope axis surfaced. Stage 2 follow_up_candidates list — IMP-29 frontend zone-level override, B4 frame_selection evidence integration, IMP-04 + IMP-05 live — remains outside this issue scope and is already documented in the carve-out doc u2 as activation gate prerequisites.)
+
+Stage 3 implementation complete pending Codex r4 verification. All 4 atomic units executed under cap. No runtime behavior change. No commit / push performed (per Stage 3 protocol — Gitea comment is the deliverable).

.orchestrator/drafts/17_stage_code-edit_codex_r1.md

@@ -0,0 +1,34 @@
+[Codex #4]
+
+Stage 3 Round #1 verification for unit u1 only.
+
+Verdict: u1 is correct and scope-locked. The route-hint comment now references IMP-17 instead of the non-existent IMP-31, the adjacent IMP-29 reject-route comment is unchanged, and the new anchor test verifies both facts. No runtime fallback code was introduced.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- git diff -- src/phase_z2_pipeline.py tests/orchestrator_unit/test_imp17_comment_anchor.py
+- Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\tests\orchestrator_unit\test_imp17_comment_anchor.py'
+- python -m pytest -q tests/orchestrator_unit/test_imp17_comment_anchor.py
+- line anchor read for src/phase_z2_pipeline.py lines 561-566
+- git diff --numstat -- src/phase_z2_pipeline.py tests/orchestrator_unit/test_imp17_comment_anchor.py
+
+Files checked:
+- D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py
+- D:\ad-hoc\kei\design_agent\tests\orchestrator_unit\test_imp17_comment_anchor.py
+
+Test results:
+- python -m pytest -q tests/orchestrator_unit/test_imp17_comment_anchor.py
+- 2 passed in 0.02s
+
+Verified facts:
+- src/phase_z2_pipeline.py line 564 contains restructure and IMP-17.
+- src/phase_z2_pipeline.py line 564 no longer contains IMP-31.
+- src/phase_z2_pipeline.py line 565 still contains reject and IMP-29.
+- _IMP05_ROUTE_HINTS runtime mapping was not changed by the diff.
+- The focused unit diff is one source comment change plus the new anchor test.
+- Existing unrelated dirty worktree entries were not touched for this verification.
+Remaining units: [u2, u3, u4]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/17_stage_code-edit_codex_r2.md

@@ -0,0 +1,51 @@
+[Codex #5]
+
+Stage 3 Round #2 verification for unit u2 only.
+
+Verdict: u2 is correct and scope-locked. The new carve-out document defines the AI fallback-only boundary, activation gate, pattern references, and Kei persona disconnect without introducing runtime fallback code or normal-path AI calls.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- git diff -- docs/architecture/IMP-17-CARVE-OUT.md
+- Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md'
+- (Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md').Count
+- line anchor reads for src/content_editor.py lines 21 and 318
+- line anchor read for src/sse_utils.py lines 16-50
+- line anchor read for src/phase_z2_pipeline.py lines 564 and 569
+- Test-Path for linked architecture docs
+- rg anchor/guardrail scan over docs/architecture/IMP-17-CARVE-OUT.md
+- git grep / Select-String / ReadAllLines checks for PHASE-Z-PIPELINE-OVERVIEW.md Step 12, Step 16, Step 17, Step 18 anchors
+- ReadAllLines check for PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md line 68
+- rg check for IMP-31 / IMP-29 in docs/architecture/IMP-17-CARVE-OUT.md
+
+Files checked:
+- D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md
+- D:\ad-hoc\kei\design_agent\src\content_editor.py
+- D:\ad-hoc\kei\design_agent\src\sse_utils.py
+- D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-AUDIT.md
+
+Test results:
+- Not run. u2 is docs-only, and the current-stage instruction says not to run full pytest unless explicitly required.
+
+Verified facts:
+- docs/architecture/IMP-17-CARVE-OUT.md exists as a new 44-line docs-only file.
+- The doc states Status = carve-out / design-only, normal-path AI calls = 0, and no runtime fallback until the activation gate clears.
+- The allowed boundary is fallback-only for Step 12 restructure and Step 16/17 retry-exhaustion paths.
+- The forbidden boundary excludes normal-path AI calls, MDX text compression, direct HTML/CSS/frame/layout/zone generation, padding/spacing/tolerance shrink, and new IMP ID allocation.
+- The activation gate is a 3-condition AND: explicit User GO, B4 frame_selection evidence integration complete, and IMP-04 + IMP-05 live.
+- Pattern references are link-only: src/content_editor.py line 21 imports stream_sse_tokens, line 318 calls stream_sse_tokens(response), and src/sse_utils.py lines 16-50 contain the SSE token parser.
+- src/phase_z2_pipeline.py line 564 references IMP-17, and line 569 emits the deterministic ai_adaptation_required route hint for restructure.
+- PHASE-Z-PIPELINE-OVERVIEW.md anchors resolve to Step 12 at line 280, Step 16 at line 318, and Step 17 at line 326.
+- PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md line 68 is the IMP-17 backlog row.
+- PHASE-Q-INSIGHT-TO-22STEP-MAP.md still has the AI repair fallback infra registry row without IMP-17 prefix; that is expected because u4 remains pending.
+- The new doc contains no IMP-31 or IMP-29 references.
+- Existing unrelated dirty worktree entries were not touched for this verification.
+Remaining units: [u3, u4]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/17_stage_code-edit_codex_r3.md

@@ -0,0 +1,39 @@
+[Codex #6]
+
+Stage 3 Round #3 verification for unit u3 only.
+
+Verdict: u3 is correct and scope-locked. The IMP-17 backlog row now links to the carve-out boundary document and carries only a short activation-gate pointer, while the binding gate definition remains in the u2 document. No runtime code was changed.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- Test-Path -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md'
+- line read for docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md lines 67-71
+- rg -n "Activation gate|User GO|B4 frame_selection|IMP-04|IMP-05|normal-path AI calls = 0|Kei persona" docs/architecture/IMP-17-CARVE-OUT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- git diff --numstat -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- column/link/guardrail/dependency checks for the IMP-17 backlog row
+- GITEA_TOKEN presence check
+
+Files checked:
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md
+
+Test results:
+- Not run. u3 is docs-only, and the current-stage instruction says not to run full pytest unless explicitly required.
+
+Verified facts:
+- The focused diff changes exactly one logical row in docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md: the IMP-17 backlog row.
+- The IMP-17 scope cell now links to [carve-out boundary + activation gate](IMP-17-CARVE-OUT.md).
+- The link target exists at docs/architecture/IMP-17-CARVE-OUT.md.
+- The row includes only a short 3-condition pointer: User GO, B4 frame_selection evidence, and IMP-04/05 live.
+- The full activation gate definition remains in docs/architecture/IMP-17-CARVE-OUT.md under "Activation gate (3-condition AND — all three required)".
+- Existing IMP-17 pattern text remains in the row: httpx + SSE streaming + retry + JSON parse pattern reference, light_edit / restructure proposal.
+- The IMP-17 guardrail column still states normal path AI calls = 0 and Kei persona disconnect.
+- The IMP-17 dependency column still states soft link: IMP-04 + IMP-05.
+- Adjacent rows IMP-16 and IMP-18 are unchanged by the diff.
+- No runtime fallback implementation or normal-path AI call was introduced.
+Remaining units: [u4]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/17_stage_code-edit_codex_r4.md

@@ -0,0 +1,38 @@
+[Codex #7]
+
+Stage 3 Round #4 verification for unit u4 only.
+
+Verdict: u4 is correct and scope-locked. The INSIGHT-MAP axis registry row now has the IMP-17 prefix and links to the carve-out document, while the Step 12/16/17 scope, G3 source, pending status, and normal-path=no fallback-only contract remain unchanged. No runtime code was changed.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- git diff --numstat -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- Test-Path -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md'
+- line read for docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md lines 120-126
+- rg -n "IMP-17 AI repair fallback infra|AI repair fallback infra|IMP-17-CARVE-OUT|no \(AI fallback only\)" docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs/architecture/IMP-17-CARVE-OUT.md
+- Test-Path -LiteralPath 'D:\ad-hoc\kei\design_agent\.orchestrator\drafts\17_stage_code-edit_codex_r4.md'
+- GITEA_TOKEN presence check
+
+Files checked:
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md
+
+Test results:
+- Not run. u4 is docs-only, and the current-stage instruction says not to run full pytest unless explicitly required.
+
+Verified facts:
+- The focused diff changes exactly one logical row in docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md.
+- git diff --numstat for the u4 file is 1 insertion and 1 deletion.
+- Line 123 now starts with "IMP-17 AI repair fallback infra" and links to IMP-17-CARVE-OUT.md.
+- The linked carve-out document exists at docs/architecture/IMP-17-CARVE-OUT.md.
+- The row still targets Step 12, 16, 17.
+- The row still references the G3 httpx + SSE streaming + retry + JSON parse pattern.
+- The row status remains pending.
+- The normal-path column remains "no (AI fallback only)".
+- Adjacent rows 122 and 124 are unchanged by the focused diff.
+- The existing Step 12 cross-reference at line 88 still contains "AI repair fallback infra axis" and continues to resolve to the registry row phrase.
+- No runtime fallback implementation or normal-path AI call was introduced.
+Remaining units: none
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_commit-push_claude_r1.md

@@ -0,0 +1,86 @@
+📌 **[오케스트레이터] Stage 5: 커밋 및 푸쉬 R1 — [Claude #13]**
+
+■ 핵심 결정 (Korean)
+- IMP-17 carve-out 4 units (u1 anchor 정정 + u2 carve-out 문서 + u3 backlog row + u4 INSIGHT-MAP row) 단일 커밋으로 묶어 `e10ec36` 로 lock 완료.
+- 스테이징 = **정확히 5 파일만** (Stage 3 plan 와 1:1 일치). 다른 모든 working-tree 변경 (orchestrator.py / test_subprocess_cleanup.py / matching/ 산출물 / chromedriver / samples / templates 등) 은 의도적으로 untouched 유지.
+- Push = `origin` (GitHub) + `slide2` (Gitea C.E.L_Slide_test2 — issue #17 host) 양쪽 모두 성공. 두 remote 의 `refs/heads/main` 모두 `e10ec36` 로 일치.
+- 런타임 행동 변경 = 0. 실행 파일 delta = `src/phase_z2_pipeline.py:564` 의 comment 한 줄 뿐. normal-path AI 호출 수 변동 없음.
+- 본 commit 으로 IMP-17 carve-out 슬롯이 design-only 상태로 영구 귀속됨. activation gate (3-cond AND) 미충족 동안 runtime fallback 코드 진입 봉인.
+
+■ 범위 제외
+- runtime fallback 구현, content_editor.py 부활, Kei persona / Kei-API endpoint 재연결, normal-path AI 호출 도입, IMP-29 frontend override 본 작업, B4 frame_selection evidence 본 작업, 신규 IMP ID 발급 — 본 commit 범위 밖.
+- working tree 의 다른 변경 (orchestrator.py / matching/ 산출물 / chromedriver / samples 등) — 별 axis 책임, 본 커밋에 비포함.
+
+■ 다음 단계
+- Stage 6 (이슈 종료 + final report). Codex 양측 검증 후 issue #17 close.
+
+=== EXIT REPORT (English, binding contract) ===
+
+**commit_sha**: `e10ec36` (full: `e10ec3661786b8510d5e4ab16291c11fc36f9648`)
+
+**parent_sha**: `23ba8b6` (HEAD at Stage 1 entry — `feat(IMP-16): U1 H3 verification utility port + U2 wiring design`)
+
+**files_committed** (5 files, +76 / −3):
+- `src/phase_z2_pipeline.py` — +1 / −1 (line 564 comment: `IMP-31` → `IMP-17 — carve-out, AI fallback only, normal path 밖`)
+- `docs/architecture/IMP-17-CARVE-OUT.md` — new, +44 lines
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — +1 / −1 (row 68 IMP-17 cell)
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` — +1 / −1 (row 123 prefix + carve-out link)
+- `tests/orchestrator_unit/test_imp17_comment_anchor.py` — new, +29 lines (2 anchor tests)
+
+**files_excluded** (intentionally NOT staged — out of IMP-17 scope, untouched in working tree):
+- `orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py` — separate axis.
+- `tests/matching/*`, `chromedriver/*`, `samples/*`, `templates/catalog/*`, `templates/phase_z2/families/*`, `templates/phase_z2/frames/*`, `scripts/*`, `tests/PIPELINE.md` etc. — separate axes / pre-existing untracked artifacts.
+- `.orchestrator/`, `.claude/settings.local.json` — local-only.
+
+**push_targets**:
+- `origin` (GitHub `keimin86/design_agent`) → `refs/heads/main` = `e10ec36` ✅
+- `slide2` (Gitea `Kyeongmin/C.E.L_Slide_test2` — issue #17 host) → `refs/heads/main` = `e10ec36` ✅
+- `gitea` (Gitea `Kyeongmin/_Geulbeot`) — NOT pushed (separate project per memory `reference_design_agent_remotes`).
+
+**verification**:
+- `git ls-remote origin refs/heads/main` = `e10ec36...` ✓
+- `git ls-remote slide2 refs/heads/main` = `e10ec36...` ✓
+- `git rev-parse HEAD` = `e10ec36...` ✓
+- `git diff --staged --stat` pre-commit = 5 files, +76/−3 (matches Stage 3 plan exactly).
+- `python -m pytest tests/orchestrator_unit/test_imp17_comment_anchor.py -v` = 2 passed in 0.03s (pre-commit re-run).
+
+**commit_message_summary**: Conventional `feat(IMP-17):` prefix with carve-out + 3-cond AND gate headline, per-unit (u1/u2/u3/u4) breakdown in body, anchor test note, runtime-impact=0 declaration, `Refs: gitea #17`, Co-Authored-By trailer.
+
+**guardrails (verified intact on commit)**:
+- normal_path_ai_calls: 0 (no runtime code added; only comment delta).
+- kei_persona_decoupled: yes (carve-out doc §AI 격리 + Kei persona 단절).
+- no_mdx_compression: yes (Phase Z spacing direction guardrail honored).
+- imp29_line_565: untouched (anchor test enforces).
+- new_imp_id_minted: no (carve-out absorbed into IMP-17 slot).
+- pattern_reference_mode: link-only (content_editor.py + sse_utils.py — Archive Candidate, no port).
+- scope_lock: 5 files only; no scope bleed to working-tree concurrents.
+
+**unresolved_questions**: none. Stage 1 Q1/Q2/Q3 locked in Stage 2 R2; Stage 4 only flagged Step 12 anchor line-number drift in IMP-17-CARVE-OUT.md (lines 280-287 reflect Stage 2 plan snapshot; section anchors still resolve by heading — out-of-IMP-17-scope docs sync.)
+
+**evidence**:
+- `git log -1 --stat` confirms commit `e10ec36` author kyeongmin, 5 files, +76/−3.
+- `git ls-remote origin refs/heads/main` = `e10ec3661786b8510d5e4ab16291c11fc36f9648 refs/heads/main`.
+- `git ls-remote slide2 refs/heads/main` = `e10ec3661786b8510d5e4ab16291c11fc36f9648 refs/heads/main`.
+- `git push origin main` output: `23ba8b6..e10ec36  main -> main` (GitHub).
+- `git push slide2 main` output: `23ba8b6..e10ec36  main -> main` (Gitea).
+- `pytest tests/orchestrator_unit/test_imp17_comment_anchor.py` = 2/2 PASS.
+
+**source_comment_ids**: Claude #13 (Stage 5 R1 — this comment).
+
+rewind_target: (n/a — push complete)
+
+FINAL_CONSENSUS: YES
+
+=== EVIDENCE ===
+- commit_sha: `e10ec3661786b8510d5e4ab16291c11fc36f9648` (HEAD)
+- parent_sha: `23ba8b6` (Stage 1 entry)
+- files staged + committed (5, +76/−3): src/phase_z2_pipeline.py (line 564 comment), docs/architecture/IMP-17-CARVE-OUT.md (new, 44L), docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md (row 68), docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md (row 123), tests/orchestrator_unit/test_imp17_comment_anchor.py (new, 29L).
+- pre-commit `git diff --staged --stat` = exactly 5 files, +76/−3 (matches Stage 3 plan).
+- pre-commit `pytest tests/orchestrator_unit/test_imp17_comment_anchor.py` = 2 passed in 0.03s.
+- `git push origin main` = `23ba8b6..e10ec36  main -> main` (GitHub `keimin86/design_agent`).
+- `git push slide2 main` = `23ba8b6..e10ec36  main -> main` (Gitea `Kyeongmin/C.E.L_Slide_test2` — issue #17 host).
+- post-push `git ls-remote origin refs/heads/main` = `e10ec36...` ✓ matches HEAD.
+- post-push `git ls-remote slide2 refs/heads/main` = `e10ec36...` ✓ matches HEAD.
+- runtime behavior change: 0 (only comment delta in executable file; normal-path AI calls remain 0).
+- excluded from staging (working tree untouched): orchestrator.py, tests/orchestrator_unit/test_subprocess_cleanup.py, tests/matching/*, chromedriver/*, samples/*, templates/*, scripts/*, .orchestrator/, .claude/settings.local.json.
+- guardrails verified: AI 격리 contract, Kei persona disconnect, no MDX compression, IMP-29 line 565 preserved, no new IMP ID minted, pattern reference link-only.

.orchestrator/drafts/17_stage_commit-push_codex_r1.md

@@ -0,0 +1,55 @@
+[Codex #14]
+
+Stage 5 commit/push verification for IMP-17.
+
+Verdict: PASS. Local HEAD and both pushed remotes point to the same IMP-17 commit, and the commit contains only the five planned IMP-17 files.
+
+Commit checked:
+- SHA: e10ec3661786b8510d5e4ab16291c11fc36f9648
+- Parent: 23ba8b68cdc315cbab542ce8c285caf2f431dbe5
+- Subject: feat(IMP-17): AI repair fallback infra carve-out - design-only boundary + 3-cond AND gate
+
+Remote reflection:
+- origin refs/heads/main = e10ec3661786b8510d5e4ab16291c11fc36f9648
+- slide2 refs/heads/main = e10ec3661786b8510d5e4ab16291c11fc36f9648
+
+Committed file scope:
+- docs/architecture/IMP-17-CARVE-OUT.md (+44)
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md (+1/-1)
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md (+1/-1)
+- src/phase_z2_pipeline.py (+1/-1)
+- tests/orchestrator_unit/test_imp17_comment_anchor.py (+29)
+
+Unintended files in commit: none found. `git diff-tree --no-commit-id --name-only -r HEAD` matches the Stage 3/4 planned 5-file scope exactly.
+
+Working tree note: unrelated modified/untracked files remain in the local working tree (`orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py`, `.orchestrator/`, `.claude/settings.local.json`, `chromedriver/`, `samples/`, `scripts/`, `templates/`, and `tests/matching/` artifacts), but they are not included in commit e10ec36.
+
+Runtime behavior check: PASS by commit scope. The only executable-file delta in this commit is the `src/phase_z2_pipeline.py` line 564 comment change; no runtime fallback implementation or normal-path AI invocation was added.
+
+=== EVIDENCE ===
+Commands run:
+- `git rev-parse HEAD`
+- `git show --name-status --oneline --stat --no-renames HEAD`
+- `git ls-remote origin refs/heads/main`
+- `git ls-remote slide2 refs/heads/main`
+- `git show -s --format='%H%n%P%n%s' HEAD`
+- `git diff-tree --no-commit-id --name-only -r HEAD`
+- `git diff-tree --no-commit-id --numstat -r HEAD`
+- `git status --short`
+Files checked:
+- docs/architecture/IMP-17-CARVE-OUT.md
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- src/phase_z2_pipeline.py
+- tests/orchestrator_unit/test_imp17_comment_anchor.py
+Commit SHA: e10ec3661786b8510d5e4ab16291c11fc36f9648
+Verified facts:
+- Local HEAD equals e10ec3661786b8510d5e4ab16291c11fc36f9648.
+- origin/main equals e10ec3661786b8510d5e4ab16291c11fc36f9648.
+- slide2/main equals e10ec3661786b8510d5e4ab16291c11fc36f9648.
+- Commit parent is 23ba8b68cdc315cbab542ce8c285caf2f431dbe5.
+- Commit file set is exactly the planned five IMP-17 files.
+- Commit numstat is +76/-3 across those five files.
+- No unrelated working-tree files were included in HEAD.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_final-close_claude_r1.md

@@ -0,0 +1,103 @@
+📌 **[오케스트레이터] Stage 6: 최종 확인 / close — [Claude #1]**
+
+■ **핵심 결정**
+- IMP-17 **carve-out anchor 목표 달성 → close 권고**. 본 issue 는 design-only carve-out (boundary + 3-cond AND gate + pattern link-only + AI 격리 contract) 이며 runtime fallback 코드 task 가 아님. 목표 4 항목 전부 commit `e10ec36` 에 landed.
+- Commit `e10ec3661786b8510d5e4ab16291c11fc36f9648` 가 **origin/main 과 slide2/main 양쪽에서 verbatim 일치** (`git ls-remote` 재확인). 5-file scope (src 1 / docs 3 / test 1) 일치, 의도 외 파일 0.
+- Runtime behavior delta = 0. `src/phase_z2_pipeline.py:564` 한 줄 comment 만 코드 변경; line 565 IMP-29 reference 무손상. IMP-31 ghost 는 src/docs 어디에도 없음 (anchor test 의 negative assertion 만 잔존, 의도).
+- 활성화는 영구 잠금 아님. **3-condition AND gate** (User GO ∧ B4 frame_selection evidence ∧ IMP-04 catalog + IMP-05 V4 fallback live) 충족 시 별도 runtime issue 분기. 본 issue 가 후속 runtime 까지 떠안지 않음.
+- 신규 follow-up issue 미발급. 활성화 게이트의 3 prerequisite 는 모두 기존 backlog 안에 존재 (B4 = frame_selection evidence axis, IMP-04 = catalog 확장, IMP-05 = V4 fallback). carve-out 이 그 link 만 보유.
+
+■ **goal vs result (issue body 대조)**
+
+| issue body scope | delivered | location |
+|---|---|---|
+| httpx + SSE + retry + JSON parse pattern reference | ✓ link-only | `IMP-17-CARVE-OUT.md` §Pattern shape reference |
+| Step 12 light_edit / restructure AI fallback proposal slot | ✓ boundary 정의 | `IMP-17-CARVE-OUT.md` §Allowed |
+| Step 16 / 17 retry path AI fallback (future) | ✓ boundary 정의 | `IMP-17-CARVE-OUT.md` §Allowed |
+| normal path AI 호출 0 (guardrail) | ✓ contract 명문화 | `IMP-17-CARVE-OUT.md` §Forbidden + §AI 격리 contract |
+| Kei persona 단절 (Phase Q 자산과 단절) | ✓ contract 명문화 | `IMP-17-CARVE-OUT.md` §AI 격리 + Kei persona 단절 |
+| dependency: soft link IMP-04 + IMP-05 | ✓ 3-cond AND gate 의 #3 조건으로 명시 | `IMP-17-CARVE-OUT.md` §Activation gate |
+| cross-ref: backlog / INSIGHT-MAP / pipeline overview | ✓ 모든 cross-ref 양방향 일치 | u3 backlog row + u4 INSIGHT-MAP row + carve-out doc source anchors |
+
+issue body 의 모든 axis 가 carve-out 문서로 흡수. review_loop 체크박스는 close 시 운영 측에서 일괄 ✓ 처리 권장.
+
+■ **범위 제외 (이 issue 가 떠안지 않는 것)**
+- Runtime fallback 실제 구현 (httpx + SSE + retry runtime). 3-cond AND gate 충족 시 별도 issue.
+- B4 frame_selection evidence integration 본 작업 (Step 9 axis, 별도).
+- IMP-04 catalog 확장, IMP-05 V4 fallback 활성화 (각 axis 별도).
+- IMP-29 frontend zone-level override (`:565`, 별도 carve-out 후보, 본 commit 무손상).
+- working tree 의 unrelated 변경 (orchestrator.py, samples/, scripts/, tests/matching/ 등) — 본 carve-out 범위 외.
+
+■ **다음 단계**
+- **Close issue #17** (carve-out anchor 완료). 운영 측에서 review_loop 체크박스 일괄 ✓ + carve-out / design-only / phase-z 라벨 (없으면) 추가 권장.
+- runtime 활성화는 3-cond AND gate 충족 시 신규 issue 로 분기. 본 carve-out 은 그 시점의 design 기준선으로 역할 종료.
+
+```yaml
+=== EVIDENCE ===
+commit:
+  sha: e10ec3661786b8510d5e4ab16291c11fc36f9648
+  parent: 23ba8b68cdc315cbab542ce8c285caf2f431dbe5
+  subject: "feat(IMP-17): AI repair fallback infra carve-out — design-only boundary + 3-cond AND gate"
+  author: kyeongmin <b24009@hanmaceng.co.kr>
+
+remote_verification:
+  - command: git rev-parse HEAD
+    result: e10ec3661786b8510d5e4ab16291c11fc36f9648
+  - command: git ls-remote origin refs/heads/main
+    result: e10ec3661786b8510d5e4ab16291c11fc36f9648	refs/heads/main
+  - command: git ls-remote slide2 refs/heads/main
+    result: e10ec3661786b8510d5e4ab16291c11fc36f9648	refs/heads/main
+  - parity: origin == slide2 == local HEAD (PASS)
+
+file_scope (git diff-tree --no-commit-id --numstat -r HEAD):
+  - docs/architecture/IMP-17-CARVE-OUT.md: +44 / -0 (new)
+  - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md: +1 / -1
+  - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md: +1 / -1
+  - src/phase_z2_pipeline.py: +1 / -1 (line 564 comment only)
+  - tests/orchestrator_unit/test_imp17_comment_anchor.py: +29 / -0 (new)
+  - total: +76 / -3, 5 files, 0 unintended files
+
+test_verification:
+  - command: python -m pytest tests/orchestrator_unit/test_imp17_comment_anchor.py -q
+    result: 2 passed in 0.02s
+
+ghost_reference_check:
+  - grep "IMP-31" in src/: 0 hits
+  - grep "IMP-31" in docs/: 0 hits
+  - grep "IMP-31" in tests/orchestrator_unit/test_imp17_comment_anchor.py: 2 hits (both negative-assertion anchors — intentional)
+  - .orchestrator/ orchestration drafts contain historical IMP-31 references (out of scope, append-only history)
+
+source_artifact_verification (manual Read):
+  - src/phase_z2_pipeline.py:564 → "restructure  → AI-assisted frame-aware adaptation (deferred to IMP-17 — carve-out, AI fallback only, normal path 밖)" (matches Stage 3 plan verbatim)
+  - src/phase_z2_pipeline.py:565 → "reject       → design reference only (deferred to IMP-29 frontend override)" (untouched, IMP-29 preserved)
+  - docs/architecture/IMP-17-CARVE-OUT.md → 44 lines, sections: Source anchors / Carve-out boundary (Allowed + Forbidden) / Activation gate (3-cond AND, all three) / Pattern shape reference (link-only) / AI 격리 + Kei persona 단절 contract
+
+goal_vs_result_summary:
+  - issue body axes (A1 pattern ref / A2 Step 12 fallback / A3 Step 16-17 fallback / guardrails / dependency / cross-ref): all 7 axes delivered as design-only carve-out artifacts
+  - runtime behavior change: 0 (verified)
+  - 3-cond AND gate: User GO ∧ B4 frame_selection evidence ∧ IMP-04 + IMP-05 live (verbatim in carve-out doc, all three required)
+  - AI isolation contract: normal path AI calls = 0 preserved, fallback only allowed
+  - Kei persona disconnect: EDITOR_PROMPT + Kei-API endpoint marked 영구 단절
+
+close_decision: YES — carve-out anchor goal met; runtime work explicitly deferred behind activation gate, no follow-up issue needed at this time.
+
+label_recommendations: carve-out, design-only, phase-z (운영 측 라벨 정책에 따라 적용)
+
+guardrails_intact:
+  - PZ-1 (AI=0 normal path): preserved (only comment + docs + test touched)
+  - PZ-2 (1 turn = 1 step): Stage 6 = close decision only
+  - PZ-4 (no silent shrink): no MDX compression introduced
+  - AI 격리 contract: explicit in carve-out §AI 격리 contract
+  - Kei persona disconnect: explicit in carve-out §AI 격리 + Kei persona 단절
+  - No new IMP ID minted: carve-out absorbed into IMP-17 slot
+  - Anchor sync (RULE 13): backlog / INSIGHT-MAP / carve-out doc cross-references all aligned
+
+unresolved_questions:
+  - Step 12 anchor line numbers in carve-out doc (280-287) reflect Stage 2 snapshot of PHASE-Z-PIPELINE-OVERVIEW.md; section headings still resolve, but line numbers may drift. Refresh in a future docs-sync pass (out of IMP-17 scope, noted at Stage 4).
+
+source_comment_ids:
+  - "[Claude #11] Stage 4 verify PASS"
+  - "[Codex #12] Stage 4 independent verify PASS"
+  - "[Codex #14] Stage 5 commit/push PASS"
+  - "[Claude #1] Stage 6 final-close (this comment)"
+```

.orchestrator/drafts/17_stage_final-close_codex_r1.md

@@ -0,0 +1,62 @@
+[Codex #15]
+
+Stage 6 final independent verification result: PASS.
+
+Status=3-axis:
+- Scope: PASS. Commit e10ec3661786b8510d5e4ab16291c11fc36f9648 contains exactly the planned IMP-17 carve-out files: 1 source comment, 3 docs files, and 1 anchor test.
+- Regression: PASS. Runtime behavior delta is zero; the only executable-file delta is the route-hint comment at src/phase_z2_pipeline.py:564. No normal-path AI call was added.
+- Close readiness: PASS. The issue-body axes are represented as design-only carve-out boundaries, link-only pattern references, and the 3-condition AND activation gate. Close OK.
+
+Verification notes:
+- origin/main and slide2/main both resolve to e10ec3661786b8510d5e4ab16291c11fc36f9648.
+- src/phase_z2_pipeline.py:564 now points restructure to IMP-17 carve-out, AI fallback only, normal path out.
+- src/phase_z2_pipeline.py:565 still points reject to IMP-29 frontend override.
+- docs/architecture/IMP-17-CARVE-OUT.md defines design-only status, normal-path AI calls = 0, Step 12/16/17 fallback boundaries, link-only pattern refs, Kei persona disconnect, and the 3-condition AND activation gate: User GO + B4 frame_selection evidence + IMP-04/IMP-05 live.
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68 links the IMP-17 carve-out doc and keeps the gate pointer.
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123 prefixes the AI repair fallback infra registry row with IMP-17 and preserves normal path = no / AI fallback only.
+- The old IMP-31 route hint is not present in src or docs. The only remaining IMP-31 mentions found are in tests/orchestrator_unit/test_imp17_comment_anchor.py as historical context and a negative assertion.
+- The working tree has unrelated pending local changes; they are not part of commit e10ec36 and do not affect the pushed IMP-17 scope verification.
+
+=== EVIDENCE ===
+Commands run:
+- git rev-parse HEAD
+- git show -s --format='%H%n%P%n%s%n%an <%ae>' HEAD
+- git diff-tree --no-commit-id --name-only -r HEAD
+- git diff-tree --no-commit-id --numstat -r HEAD
+- git status --short
+- git ls-remote origin refs/heads/main
+- git ls-remote slide2 refs/heads/main
+- Select-String -Path 'src\\phase_z2_pipeline.py' -Pattern 'IMP-17|IMP-29|IMP-31' -Context 1,1
+- Select-String -Path 'docs\\architecture\\IMP-17-CARVE-OUT.md' -Pattern 'Activation gate|User GO|B4 frame_selection|IMP-04|IMP-05|Pattern shape reference|content_editor.py|sse_utils.py|normal path|Kei' -Context 1,1
+- Select-String -Path 'docs\\architecture\\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md','docs\\architecture\\PHASE-Q-INSIGHT-TO-22STEP-MAP.md' -Pattern 'IMP-17|AI repair fallback infra|carve-out|normal_path|normal path' -Context 0,0
+- Select-String -Path 'tests\\orchestrator_unit\\test_imp17_comment_anchor.py' -Pattern 'IMP-17|IMP-31|IMP-29|line' -Context 1,1
+- rg -n "IMP-31" src docs tests/orchestrator_unit/test_imp17_comment_anchor.py
+- git show --unified=0 -- src/phase_z2_pipeline.py
+- python -m pytest tests/orchestrator_unit/test_imp17_comment_anchor.py -q
+- git log -1 --name-status --format='%H%n%s'
+
+Files checked:
+- D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py
+- D:\ad-hoc\kei\design_agent\docs\architecture\IMP-17-CARVE-OUT.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- D:\ad-hoc\kei\design_agent\tests\orchestrator_unit\test_imp17_comment_anchor.py
+
+Test results:
+- python -m pytest tests/orchestrator_unit/test_imp17_comment_anchor.py -q: PASS, 2 passed in 0.02s.
+- Full pytest was not run because Stage 6 did not explicitly require it.
+
+Commit SHA:
+- e10ec3661786b8510d5e4ab16291c11fc36f9648
+
+Verified facts:
+- HEAD is e10ec3661786b8510d5e4ab16291c11fc36f9648, parent 23ba8b68cdc315cbab542ce8c285caf2f431dbe5.
+- Commit subject is feat(IMP-17): AI repair fallback infra carve-out - design-only boundary + 3-cond AND gate.
+- origin/main and slide2/main both point to e10ec3661786b8510d5e4ab16291c11fc36f9648.
+- Commit file scope is exactly: docs/architecture/IMP-17-CARVE-OUT.md; docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md; docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md; src/phase_z2_pipeline.py; tests/orchestrator_unit/test_imp17_comment_anchor.py.
+- Numstat is +44 new carve-out doc, +29 new anchor test, and three +1/-1 modifications.
+- src/phase_z2_pipeline.py executable delta is one comment line only.
+- IMP-29 line 565 remains preserved.
+- IMP-17 activation remains gated; runtime fallback implementation is not landed.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_problem-review_claude_r1.md

@@ -0,0 +1,149 @@
+[Claude #1] Stage 1 — IMP-17 problem review (AI repair fallback infra carve-out)
+
+### Root-cause framing
+
+IMP-17 is a **carve-out axis** — explicitly flagged "normal path 밖 (AI fallback only)" in both the Gitea issue body and the source maps. It is not a normal-path implementation task; it is the **axis registry slot** that holds the *infrastructure pattern reference* for a future AI repair fallback runtime. Phase Z `phase_z2_pipeline.py` currently labels `restructure` V4 outputs as `ai_adaptation_required` (line 569) with an inline comment "deferred to IMP-31" (line 564) — a forward-reference ID that does **not exist** in the backlog (which ends at IMP-28). IMP-17 (Gitea #17) is the actual carve-out slot. The numeric mismatch needs decision in this cycle (correct comment vs. accept dual numbering).
+
+The Phase Q file `src/content_editor.py` already contains the pattern source — `httpx` + SSE streaming + retry + JSON parse — but the entire file is **Archive Candidate** (PHASE-Q-AUDIT.md §2.6 line 660–663; G3 footnote line 681). The only salvageable artifact is the *infrastructure pattern shape*, not the code: Kei persona prompts (`EDITOR_PROMPT`), Kei-API endpoint config, and Kei-specific retry semantics are **all archived** per AI 격리 contract (memory: feedback_ai_isolation_contract) and Phase Q persona disconnect rule (PHASE-Q-AUDIT.md §2.6 + the issue body's guardrail "Kei persona 단절").
+
+IMP-17 scope is therefore *not a port*. It is **a carve-out boundary lock** — recording (a) what the AI fallback runtime will eventually look like, (b) what it must never touch, (c) which Phase Q artifacts remain Archive and which pattern is referenced, (d) what activation gate must fire before any code lands. Code implementation is **gated behind axis activation** which is currently "(별 axis priority — pending)" per backlog row 68.
+
+### Verified facts (value + path + upstream)
+
+- IMP-17 backlog row = `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68` — priority `(별 axis priority — pending)`, soft link `IMP-04 + IMP-05`, guardrail `normal path AI 호출 0 — 본 axis = fallback only, normal path 와 분리 설계 / Kei persona 단절 (Phase Q 자산과 단절)`
+- Axis registry source = `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123` (§3 row "AI repair fallback infra | Step 12, 16, 17 | §2.6 G3 | pending | **no (AI fallback only)**") — the only §3 row with `normal path 여부 = no` (all 8 other axes are `yes`). Confirms unique carve-out status.
+- Phase Q origin = `docs/architecture/PHASE-Q-AUDIT.md:681` (§2.6 G3 footnote): "AI repair fallback (Phase Z step12 light_edit/restructure) 설계 axis 활성 시, content_editor 의 *infrastructure pattern* (`httpx` + SSE streaming + retry + JSON parse) 이 reference 가능. 단 §0-B Audit 범위 lock (L3) 따라 *별 axis 활성 시 새 기준으로 재검토*"
+- Pipeline overview cross-ref = `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md`:
+  - line 281–283 (Step 12 Slot Payload): "*deterministic mapper* 가 기본. *AI 는 normal path 에 없음*. AI 가능 위치 (제한적): light_edit / restructure 에서 content_object → slot 배치 proposal 필요 시"
+  - line 284 (AI 금지): "MDX 원문 요약·삭제 / HTML·CSS 직접 생성 / 새 디자인 임의 / layout·frame 임의 선택" — this is the AI output schema fence
+  - line 287 (Gap): "restructure label 의 AI proposal path 미구현 (현재 restructure 는 filter)" — confirms restructure-as-filter today, restructure-as-AI-proposal in future
+- Phase Q infra files in repo (unchanged):
+  - `src/kei_client.py` (1471 lines, Kei API + httpx + SSE) — Phase Q, Archive Candidate (G3 file scope)
+  - `src/sse_utils.py` (50 lines, SSE parsing helpers) — Phase Q
+  - `src/content_editor.py` (475 lines, `_call_kei_editor_with_retry` + `EDITOR_PROMPT` + `fill_content`) — Phase Q, Archive Candidate
+- Phase Z isolation (verified now): `grep -n "kei_client\|sse_utils\|content_editor" src/phase_z2_pipeline.py` → 0 hits. **Phase Z pipeline has zero coupling to Phase Q AI infra.** No refactor risk.
+- Phase Q AI infra import sites (verified): `grep -rln "from src.kei_client\|import kei_client\|from src.content_editor\|import content_editor\|from src.sse_utils\|import sse_utils" --include="*.py"` → 4 hits, all Phase Q (`src/pipeline.py`, `src/kei_client.py` self, `src/design_director.py`, `src/content_editor.py` self). Phase Z is completely outside.
+- Forward-reference anomaly (in-scope catch): `src/phase_z2_pipeline.py:564` source comment says "(deferred to IMP-31)" — IMP-31 does NOT exist in `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (last ID = IMP-28; same scan via `grep -n "IMP-29\|IMP-30\|IMP-31\|IMP-32" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → 0 hits). The same file also references "IMP-29 frontend override" (line 565). Both IDs are forward-references to unwritten backlog axes. IMP-17 (Gitea issue #17) is the **actual** AI repair fallback carve-out slot. The author of the source comment did not align the source ID with the active backlog ID.
+- Soft dependency status:
+  - IMP-04 (catalog 32-frame expansion): in-flight per git log — `git log --all --oneline | grep -iE "IMP-04"` → 16+ commits (Track A 1/16 through 8/16, F12/F14/F17/F18 follow-ups). Active landing.
+  - IMP-05 (V4 fallback rank-2/3 + status semantics): partial — commits `15c5b9a` (initial deterministic bridge), `21476ae` (complete fallback evidence + dedup), `23d1b25` (Step 9 candidate evidence guard). Has shipped baseline.
+  - **Both soft deps have meaningful runtime today.** IMP-17 activation surface (catalog candidates + `PASS_WITH_FALLBACK` semantics) exists in repo. But activation gate (priority pending) still blocks.
+- HEAD = `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`; branch = `main`; working tree dirty on orchestrator infra (`orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py`, `.orchestrator/`, `chromedriver/`, …) — unrelated to IMP-17.
+
+### Scope-lock
+
+IMP-17 is the only §3 axis registry row marked `normal path 여부 = no`. Two stage-1 scope-lock interpretations are coherent:
+
+**Interpretation A — Code-now (Unit A + Unit B deferred)**: write a dormant Phase Z infra stub (httpx-async + SSE + retry + JSON parse helpers), unwired to runtime, with unit tests; defer call-site wiring to a later cycle. **Risk**: dormant code drift — module ages without activation. Also: writing a fresh Anthropic-direct AI client without an active call site is premature design (which model? which prompt schema? which output schema?) and likely needs rewrite when activation actually fires.
+
+**Interpretation B — Docs-only carve-out (RECOMMENDED)**: lock the carve-out *contract* as a new design doc + a small in-cycle anchor fix; no Phase Z code in this cycle. Aligns with "(별 axis priority — pending)" and the issue body's verb choice "pattern reference" (not "pattern port"). Aligns with PZ-1 (AI=0 normal path) and feedback_no_hardcoding (process > code).
+
+**SCOPE-LOCKED (RECOMMENDED — Interpretation B, docs-only):**
+
+In this cycle:
+
+1. New doc `docs/architecture/PHASE-Z-AI-REPAIR-FALLBACK-CARVEOUT.md` — sole authoritative spec for what the future AI fallback runtime will look like, with these sections:
+   - §1 Boundary: list of normal-path code paths that MUST NOT call AI fallback (enumerate `phase_z2_pipeline.py` entry points + Step 12/13 normal flow)
+   - §2 Activation gate: explicit GO conditions (IMP-04 catalog gate + IMP-05 fallback gate + user GO + axis priority lift)
+   - §3 Pattern reference: pointer to Phase Q `src/content_editor.py` (`_call_kei_editor_with_retry`) + `src/kei_client.py` + `src/sse_utils.py` — *reference shape only*, NOT a port target. Explicit annotation: "vendor TBD", "prompt schema TBD", "Kei persona excluded"
+   - §4 Phase Z target API surface (DRAFT — not implemented): proposed `src/phase_z2_ai_fallback/` package boundary, naming conventions, separation invariants (`fallback_` prefix or package isolation)
+   - §5 Output schema contract: AI fallback outputs are **Internal Region placement proposals** or **Frame Slot remapping proposals** or **restructure proposals**, NEVER raw HTML / CSS / MDX rewrites (per pipeline overview line 284)
+   - §6 Failure semantics: how AI fallback failure propagates to `Step 20 Slide Status` (reuse `PASS_WITH_FALLBACK` from IMP-05 vs new enum — open question)
+   - §7 Trace/debug.json schema additions (additive only, schema versioned)
+   - §8 Cross-refs: PHASE-Q-AUDIT.md §2.6 G3, PHASE-Q-INSIGHT-TO-22STEP-MAP.md §3, PHASE-Z-PIPELINE-OVERVIEW.md Step 12/16/17, PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md §2 row IMP-17
+
+2. Anchor-fix `src/phase_z2_pipeline.py:564` — single-line edit to replace "(deferred to IMP-31)" with the correct issue reference: "(deferred to IMP-17 — see PHASE-Z-AI-REPAIR-FALLBACK-CARVEOUT.md)". Same anchor-fix evaluation for line 565 (`IMP-29 frontend override`) — out of IMP-17 scope, flag for separate backlog catch.
+
+3. Optionally: backlog row 68 update to add status `scope-locked` (per workflow §1).
+
+**OUT OF SCOPE (Unit B — runtime wiring deferred):**
+
+- No new Phase Z module under `src/phase_z2_ai_fallback/` or similar. No code skeleton.
+- No httpx async client. No SSE streaming utility. No retry policy code. No JSON parse helpers.
+- No Step 12 restructure AI proposal trigger. No Step 17 AI fallback action runner.
+- No live AI invocation anywhere in Phase Z runtime.
+
+Activation gate (all must fire before Unit B opens as a separate issue):
+- IMP-04 catalog expansion verified complete (Track A 16/16 frame activations + matrix coverage)
+- IMP-05 V4 fallback `PASS_WITH_FALLBACK` semantics verified complete
+- User explicit GO on axis activation
+- New issue (likely IMP-17-B or follow-up) for runtime wiring
+
+**OUT OF SCOPE (Phase Q file mutation — G3 isolation):**
+
+- `src/content_editor.py` untouched. Phase Q `pipeline.py` chain keeps using it. IMP-17 = doc-level reference, not code mutation.
+- `src/kei_client.py` untouched. Same isolation.
+- `src/sse_utils.py` untouched.
+
+**OUT OF SCOPE (axis bleed):**
+
+- IMP-29 frontend override (line 565 forward-ref) — not in active backlog, separate catch.
+- IMP-31 numeric placeholder in pipeline.py:564 — fix to IMP-17 reference only; do not invent IMP-31 entry in backlog.
+
+### Guardrails
+
+- G1 (PZ-1, RULE 0, issue body invariant): Normal path AI call count = 0. Verification: `grep -rln "kei_client\|sse_utils\|content_editor\|anthropic\|httpx" src/phase_z2_*.py` must remain 0-hit after this cycle's edits. The only `httpx`/`anthropic`/`kei_client` references in repo stay confined to Phase Q files.
+- G2 (RULE 7, no hardcoding): MDX 03/04/05 not referenced in the carve-out doc. Doc is content-agnostic by construction.
+- G3 (Phase Q isolation): No edits to `src/content_editor.py`, `src/kei_client.py`, `src/sse_utils.py`, `src/pipeline.py`, `src/design_director.py`. Phase Q archive remains untouched.
+- G4 (Kei persona disconnect — issue body invariant): The carve-out doc explicitly excludes Kei persona reuse — no `EDITOR_PROMPT` reference as a valid pattern, no Kei-API endpoint as a valid endpoint, no Kei response format as valid schema. All three are listed as **excluded references** in §3 of the new doc.
+- G5 (axis activation gate): Doc §2 enumerates the gate conditions and explicitly states "no code in this cycle".
+- G6 (PZ-3, no speculative): Doc §4 / §5 / §6 fields marked "TBD" or "DRAFT" must NOT contain hypothetical vendor names, hypothetical model IDs, hypothetical prompt templates, or hypothetical JSON schemas beyond the *invariant rules* (output shape = content-unit / Internal Region / Frame Slot proposal — known from PHASE-Z-PIPELINE-OVERVIEW.md line 284).
+- G7 (RULE 13 anchor sync): Commit-push stage updates BOTH `docs/architecture/PHASE-Z-AI-REPAIR-FALLBACK-CARVEOUT.md` (new) AND `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` row 68 (status field) AND `src/phase_z2_pipeline.py:564` comment in the **same commit**. INSIGHT-MAP §3 row 123 reference stays "pending" (axis not activated — only carve-out boundary locked). Anchor SHA discipline.
+- G8 (PZ-2, 1 turn = 1 step): Stage 1 closes here. Stage 2 = simulation plan only; no jumping ahead to write the doc body. No mention of Stage 3 timing or Unit B activation timing.
+- G9 (workflow atomicity, 1 commit = 1 decision unit): commit = new carve-out doc + anchor-fix in pipeline.py line 564 + backlog row 68 status update — all 3 are one decision unit (carve-out boundary lock). Existing working-tree WIP (orchestrator.py etc.) MUST NOT be staged under this issue.
+- G10 (sample-passing avoidance, RULE 0): No reference to MDX 03/04/05 frames or sample fixtures in the carve-out doc. Doc applies to all 32 frames uniformly.
+
+### Risk
+
+Low (docs-only). Failure modes:
+
+- (a) Doc body accidentally describes a vendor-specific runtime (e.g., "use Anthropic Claude 4.5 with prompt X") → G6 violation; rewind to `stage_3_edit`. Doc must stay invariant-level.
+- (b) Doc accidentally proposes a Phase Z module path under `src/phase_z2_ai_fallback/` as if it exists → G5 violation (premature module manifestation); rewind to `stage_3_edit`. All module paths in §4 must be marked "DRAFT — not implemented".
+- (c) Anchor-fix touches more than line 564 of `phase_z2_pipeline.py` → G3 violation (scope creep into Phase Z runtime); rewind to `stage_3_edit`.
+- (d) Working-tree WIP staged into IMP-17 commit → G9 violation; rewind to `stage_5_push`.
+- (e) Backlog row 68 update accidentally changes dependency formula (`soft link: IMP-04 + IMP-05`) or priority field → G7 anchor-sync drift; rewind to `stage_3_edit`.
+- (f) Carve-out doc accidentally re-introduces `EDITOR_PROMPT` body, Kei API endpoint URL, or Kei response format as "reference patterns to copy" → G4 violation (Kei persona contamination); rewind to `stage_3_edit`.
+
+### Open questions for Codex r1
+
+1. **Interpretation A vs B**: Recommend B (docs-only carve-out lock). Confirm vs A (write dormant Unit A code skeleton). Risk asymmetry: A risks dormant-code drift + premature design; B risks "nothing happens" (mitigated by the anchor-fix + backlog row update being real artifacts).
+2. **Doc location + name**: Proposed `docs/architecture/PHASE-Z-AI-REPAIR-FALLBACK-CARVEOUT.md` (descriptive + axis-prefixed). Alternative: `docs/architecture/PHASE-Z-CARVEOUT-AI-FALLBACK.md` (carve-out-prefixed). Defer to Codex.
+3. **Anchor-fix scope**: include line 565 (`IMP-29 frontend override`) anchor-fix in this commit, or strictly limit to line 564 (IMP-31 → IMP-17)? IMP-29 is a separate forward-ref unrelated to AI fallback — recommend strict limit to line 564.
+4. **Backlog row 68 status**: after this Stage 1 lock, mark row 68 status `scope-locked` (per workflow §1)? Or leave `pending` until Unit B activation (since the axis is still pending)? Recommend `scope-locked` for the carve-out boundary itself, leaving axis activation `pending` as a separate semantic.
+5. **PASS_WITH_FALLBACK semantics**: §6 of the new doc proposes reusing IMP-05's `PASS_WITH_FALLBACK` enum for AI repair fallback. Alternative: new enum (e.g., `PASS_AFTER_AI_REPAIR` / `AI_REPAIR_FAILED`). Recommend documenting as open question in §6, **not** deciding in this stage — decision belongs to Unit B activation cycle.
+6. **Output schema contract clarity**: §5 of the new doc lists allowed AI outputs as "Internal Region placement proposal / Frame Slot remapping proposal / restructure proposal" (per pipeline overview line 284). Confirm this exhausts the allowed outputs vs. whether additional categories (e.g., "MDX section split proposal") should be enumerated.
+
+### Stage 1 lock summary
+
+- Root cause: IMP-17 is a *carve-out* (the only §3 axis with `normal path 여부 = no`) for the eventual AI repair fallback runtime. Phase Q `content_editor.py` contains the pattern source but is Archive Candidate; Kei persona is excluded; Phase Z has zero AI coupling today and must stay that way. A source-comment forward-reference (`pipeline.py:564 → IMP-31`) does not match the active backlog and needs anchor-fix to IMP-17.
+- Scope locked: **Interpretation B (docs-only)** — new `docs/architecture/PHASE-Z-AI-REPAIR-FALLBACK-CARVEOUT.md` capturing boundary / gate / pattern reference / target API DRAFT / output schema / failure semantics / trace schema / cross-refs; anchor-fix `src/phase_z2_pipeline.py:564`; backlog row 68 status update to `scope-locked`. No Phase Z runtime code. No Phase Q file mutation.
+- Unit B (runtime wiring) explicitly deferred behind activation gate (IMP-04 + IMP-05 complete + user GO + axis priority lift).
+- Guardrails G1–G10 above.
+
+=== EVIDENCE ===
+
+- Commands run:
+  - `git rev-parse HEAD` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`
+  - `git status --porcelain` → only orchestrator WIP (`orchestrator.py`, `.orchestrator/`, `chromedriver/`, …), no `src/phase_z2_*`, no `src/content_editor*`, no `src/kei_client*`, no `src/sse_utils*` dirty
+  - `git log --all --oneline | grep -iE "IMP-04|IMP-05|IMP-17|IMP-31|repair fallback|ai_fallback|light_edit|restructure proposal"` → 23+ IMP-04 commits, 3+ IMP-05 commits, **0 IMP-17 commits, 0 IMP-31 commits, 0 ai_fallback commits** (axis still pending)
+  - `grep -rln "from src.kei_client\|import kei_client\|from src.content_editor\|import content_editor\|from src.sse_utils\|import sse_utils" --include="*.py"` → 4 hits, all Phase Q (`src/pipeline.py`, `src/kei_client.py` self, `src/design_director.py`, `src/content_editor.py` self). Phase Z `src/phase_z2_*.py` files → 0 hits
+  - `grep -n "kei_client\|sse_utils\|content_editor\|anthropic\|httpx" src/phase_z2_pipeline.py` → 0 hits (Phase Z pipeline has zero AI infra coupling)
+  - `wc -l src/kei_client.py src/sse_utils.py src/content_editor.py` → 1471 / 50 / 475 (Phase Q AI infra surface)
+  - `grep -n "IMP-29\|IMP-30\|IMP-31\|IMP-32" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → 0 hits (backlog ends at IMP-28; IMP-29/IMP-31 in `pipeline.py:564/565` are forward-refs to unwritten axes)
+  - `grep -n "light_edit\|restructure" src/phase_z2_pipeline.py` → lines 92/93/103/104/107/563/564/568/569 (V4 label dispatch + IMP-05 route hint dict)
+- Files read:
+  - `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (row 68 IMP-17, row 48–49 IMP-04/IMP-05 dependencies, workflow §1 review loop)
+  - `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (line 123 §3 axis row; line 88 Step 12 row; lines 102–103 Step 16/17 rows; §0 invariant lock)
+  - `docs/architecture/PHASE-Q-AUDIT.md` (§2.6 lines 622–688 — content_editor audit, G1/G2/G3 footnotes; §0-B Audit 범위 lock cross-ref)
+  - `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md` (Step 12 lines 280–287; Step 16 lines 318–325; Step 17 lines 326–333; AI ban list line 284)
+  - `src/phase_z2_pipeline.py` (lines 92–107 V4 label dispatch; lines 560–578 IMP-05 route hint dict + IMP-31/IMP-29 forward-ref comments)
+  - `.orchestrator/drafts/16_stage_problem-review_claude_r1.md` (precedent format check — Claude #1 problem-review structure for sibling carve-out-ish axis IMP-16)
+- Facts cross-checked:
+  - IMP-17 is the only §3 axis with `normal path 여부 = no` — unique carve-out invariant (PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123)
+  - Phase Z pipeline has zero `kei_client` / `sse_utils` / `content_editor` / `anthropic` / `httpx` imports today — no refactor risk; carve-out boundary already de facto held by repo state
+  - Phase Q `content_editor.py` is module-level Archive Candidate (PHASE-Q-AUDIT.md §2.6 line 667); G3 footnote (line 681) explicitly defers AI repair fallback infra evaluation to "별 axis 활성 시 새 기준으로 재검토" — exactly the gate IMP-17 is named after
+  - Source-comment forward-reference `pipeline.py:564 → IMP-31` does not align with active backlog (IMP-28 last); IMP-17 is the actual carve-out slot → anchor-fix candidate in this stage's scope
+  - IMP-04 + IMP-05 soft deps have shipped baselines (23+ / 3+ commits respectively) — activation surface exists; axis priority "(pending)" is the blocking gate, not dependency readiness
+  - Pipeline overview line 284 enumerates 4 hard prohibitions on AI output (MDX 원문 요약·삭제 / HTML·CSS 직접 생성 / 새 디자인 임의 / layout·frame 임의 선택) — directly informs the carve-out doc §5 output schema contract
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_problem-review_codex_r1.md

@@ -0,0 +1,55 @@
+[Codex #1] Stage 1 problem review for IMP-17
+
+Verdict: Claude #1's root-cause framing is materially correct, with two scope-lock corrections that should be carried into Stage 2.
+
+Root cause / scope lock:
+- IMP-17 is the backlog and issue slot for AI repair fallback infrastructure, not a normal-path feature. Its backlog row explicitly says "carve-out - normal path outside", "normal path AI calls 0", "fallback only", and "Kei persona disconnected" at `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68`.
+- The axis registry confirms this is the only listed axis with `normal path` = `no (AI fallback only)` at `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123`; the surrounding rows are `yes`.
+- Step 12 remains deterministic in the normal path. The overview says the mapper is the default and AI is not in the normal path at `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md:280-287`; Step 16 and Step 17 are deterministic router/action paths with missing non-AI actions at `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md:318-332`.
+- `content_editor.py` is Archive Candidate at module level. The audit says the Kei direct slot-fill flow is outside Phase Z normal path and not a Salvage Plan target at `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-AUDIT.md:667-673`.
+
+Corrections / flags:
+- The live infrastructure pattern is no longer wholly inside `src/content_editor.py`. That file owns `httpx.AsyncClient.stream`, the Kei retry loop, and `_parse_json`, but SSE token parsing is delegated through `from src.sse_utils import stream_sse_tokens` at `D:\ad-hoc\kei\design_agent\src\content_editor.py:21` and called at `D:\ad-hoc\kei\design_agent\src\content_editor.py:318`. The SSE line/event/json token parser itself is in `D:\ad-hoc\kei\design_agent\src\sse_utils.py:16-50`. Stage 2 should name the reference as `content_editor.py` + `sse_utils.py` pattern shape, not a direct port from `content_editor.py`.
+- Claude #1 correctly flagged `IMP-31` as nonexistent in the backlog, but the adjacent inline comment also references nonexistent `IMP-29`. `rg -n "IMP-29|IMP-30|IMP-31" docs src tests` returns only `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py:564-565`, while the backlog ends at `IMP-28` at `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:93`. Stage 2 should decide whether those comments are corrected to IMP-17 / an existing issue, or documented as external numbering.
+- No required architecture files were missing in this checkout. I did not read the full Gitea comment history, per the orchestrator instruction.
+
+Scope conclusion:
+- No code implementation is authorized by this stage. The correct Stage 2 plan target is a boundary/infra-reference lock for future fallback activation: normal path AI remains 0, Kei persona assets remain archive/disconnected, and any future runtime must be behind an explicit fallback activation gate.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\.orchestrator\tmp\codex_prompt.txt'`
+- `rg -n "IMP-17|AI repair fallback|normal path|fallback only|Step 12|Step 16|Step 17|content_editor|httpx|SSE|JSON parse|retry|ai_adaptation_required|IMP-31|Archive Candidate|Kei persona|EDITOR_PROMPT" docs src tests`
+- `git status --short`
+- `rg --files`
+- `Select-String -Path 'docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md' -Pattern 'IMP-17|IMP-28|IMP-31|AI repair fallback|normal path|Kei persona' -Context 2,2`
+- `Select-String -Path 'docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md' -Pattern 'AI repair fallback|normal path|Step 12|Step 16|Step 17|no \(AI fallback only\)' -Context 2,2`
+- `Select-String -Path 'docs\architecture\PHASE-Q-AUDIT.md' -Pattern 'Archive Candidate|content_editor|AI repair fallback|httpx|SSE|retry|JSON parse|Kei persona|AI.*isolation' -Context 2,2`
+- `Select-String -Path 'docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md' -Pattern 'Step 12|Step 16|Step 17|AI|normal path|light_edit|restructure|fallback' -Context 2,2`
+- `Select-String -Path 'src\phase_z2_pipeline.py' -Pattern 'IMP-31|ai_adaptation_required|restructure|deferred|route_hint' -Context 4,4`
+- `Select-String -Path 'src\content_editor.py' -Pattern 'import httpx|EDITOR_PROMPT|_call_kei_editor_with_retry|AsyncClient|stream\(|aiter_lines|json.loads|retry|attempt|except|_parse_json' -Context 2,2`
+- `Select-String -Path 'src\sse_utils.py' -Pattern 'httpx|aiter_lines|json.loads|event:|data:|error|done|token' -Context 2,2`
+- `rg -n "IMP-31" docs src tests`
+- `Select-String -Path 'src\content_editor.py' -Pattern 'stream_sse_tokens|from src.sse_utils' -Context 2,2`
+- `Select-String -Path 'docs\architecture\PHASE-Q-AUDIT.md' -Pattern '### 2.6|G1|G2|G3|Archive Candidate \(module-level\)|httpx|SSE streaming|retry \+ JSON parse|AI repair fallback infra' -Context 3,3`
+- `Select-String -Path 'docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md' -Pattern '^\| IMP-' | Select-Object -Last 5`
+- `rg -n "IMP-29|IMP-30|IMP-31" docs src tests`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\.orchestrator\tmp\codex_prompt.txt`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-AUDIT.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py`
+- `D:\ad-hoc\kei\design_agent\src\content_editor.py`
+- `D:\ad-hoc\kei\design_agent\src\sse_utils.py`
+Test results: Not run; Stage 1 explicitly says NO pytest.
+Verified facts:
+- IMP-17 value: fallback-only carve-out, normal path AI calls 0, Kei persona disconnected. Path: `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68`. Upstream: Gitea issue #17 body and backlog cross-ref.
+- Axis registry value: `AI repair fallback infra | Step 12, 16, 17 | pending | no (AI fallback only)`. Path: `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123`. Upstream: issue #17 INSIGHT-MAP cross-ref.
+- Step 12 value: deterministic mapper default, AI not in normal path; AI only possible as limited light_edit/restructure proposal. Path: `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md:280-287`. Upstream: issue #17 22-step pipeline cross-ref.
+- Phase Q audit value: `content_editor.py` is Archive Candidate module-level, with only infra pattern reference eligible for future separate-axis reconsideration. Path: `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-AUDIT.md:667-681`. Upstream: issue #17 source `INSIGHT-MAP section 3 (G3)`.
+- Live code value: `restructure` route hint is `ai_adaptation_required`, with stale comments referencing `IMP-31` and `IMP-29`. Path: `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py:564-569`. Upstream: IMP-05 route hint comments in current code.
+- Backlog numbering value: implementation backlog ends at `IMP-28`; no `IMP-29`, `IMP-30`, or `IMP-31` appears in docs/src/tests except the two inline comments in `phase_z2_pipeline.py`. Path: `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:91-93` and `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py:564-565`. Upstream: current repo grep.
+- Live pattern value: `content_editor.py` performs `httpx.AsyncClient.stream`, retry loop, and `_parse_json`; `sse_utils.py` performs SSE line/event parsing and JSON token parsing. Paths: `D:\ad-hoc\kei\design_agent\src\content_editor.py:21`, `D:\ad-hoc\kei\design_agent\src\content_editor.py:304-328`, `D:\ad-hoc\kei\design_agent\src\content_editor.py:443-475`, `D:\ad-hoc\kei\design_agent\src\sse_utils.py:16-50`. Upstream: current code.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_simulation-plan_claude_r1.md

@@ -0,0 +1,91 @@
+[Claude #2] Stage 2 — Simulation Plan R1
+
+## Axis Coverage Checklist
+
+| # | axis | covered by |
+|---|---|---|
+| A1 | `httpx` + SSE + retry + JSON parse pattern reference | u2 §Pattern Ref |
+| A2 | Step 12 light_edit / restructure proposal (AI fallback) | u2 §Boundary + u4 |
+| A3 | Step 16 / 17 retry path AI fallback (future) | u2 §Boundary + u4 |
+| S1 | `phase_z2_pipeline.py:564` IMP-31 → IMP-17 correction | u1 |
+| S2 | Carve-out boundary doc (allowed / forbidden) | u2 §Boundary |
+| S3 | Axis activation gate | u2 §Gate + u3 |
+| S4 | Pattern shape reference (link only) | u2 §Pattern Ref |
+
+3 issue-body axes + 4 Stage-1 derivatives → 4 atomic units. No partial axis.
+
+## Stage-1 Unresolved Question Resolutions
+
+- Q1 (comment style): rewrite verbatim to `IMP-17` — single source of truth, alias dropped.
+- Q2 (activation gate): 3-condition AND — (a) explicit user GO on IMP-17, (b) B4 frame_selection evidence integration complete, (c) IMP-04 catalog 확장 + IMP-05 V4 fallback live (per backlog soft-link). All required, not OR.
+- Q3 (SSE reference style): link to `src/content_editor.py:21,318` + `src/sse_utils.py:16-50` inside carve-out doc — no verbatim inline. Archive contract preserved.
+
+## === IMPLEMENTATION_UNITS ===
+
+```yaml
+- id: u1
+  summary: Correct src/phase_z2_pipeline.py:564 comment "deferred to IMP-31" → "deferred to IMP-17 (carve-out — AI fallback only, normal path 밖)".
+  files:
+    - src/phase_z2_pipeline.py
+  tests:
+    - tests/orchestrator_unit/test_imp17_comment_anchor.py
+  estimate_lines: 6
+
+- id: u2
+  summary: Create docs/architecture/IMP-17-CARVE-OUT.md — boundary (allowed/forbidden), pattern shape reference (link only), activation gate (3-condition AND), Step 12 / 16 / 17 fallback slot anchors.
+  files:
+    - docs/architecture/IMP-17-CARVE-OUT.md
+  tests: []
+  estimate_lines: 48
+
+- id: u3
+  summary: Update docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68 IMP-17 row — append link to carve-out doc + activation gate hint cell.
+  files:
+    - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+  tests: []
+  estimate_lines: 4
+
+- id: u4
+  summary: Update docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123 — anchor IMP-17 ID prefix into existing "AI repair fallback infra" axis row.
+  files:
+    - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+  tests: []
+  estimate_lines: 3
+```
+
+## Per-unit Rationale
+
+- u1: Stage-1 scope-lock item; kills forward-ref to non-existent ID. Test asserts `IMP-31` absent + `IMP-17` present at line 564. Deterministic, atomic.
+- u2: Primary artifact. §Boundary (runtime may/may-not touch), §Pattern Reference (link only — Archive contract upheld), §Gate (3-condition AND), §Step Anchors (12/16/17 fallback slots). Zero runtime code, zero persona prompt, zero httpx port.
+- u3: Backlog row gains pointer to carve-out doc + short gate hint. Long-form gate definition lives in u2.
+- u4: Pure cross-ref hygiene — ID anchor missing; "normal path 여부 = no" already present.
+
+## Out-of-scope (deferred follow-ups)
+
+- `phase_z2_pipeline.py:565` "deferred to IMP-29" — separate axis (frontend zone override). IMP-29 also absent from backlog (ends at IMP-28). **Follow-up candidate** (not modified here).
+- `content_editor.py` / `EDITOR_PROMPT` / Kei-API endpoint revival — Archive contract upheld.
+- Step 12 normal-path determinism — unchanged.
+- Allocating new IMP IDs ≥ IMP-29 — Stage 1 lock.
+- Fallback runtime implementation — gated behind 3-condition activation gate.
+
+## Side Effects
+
+u1 = single comment line, zero behavior delta (`_IMP05_ROUTE_HINTS` unchanged). u2 = new doc file only. u3/u4 = row-cell edits, no schema change. Test added in u1 = doc-anchor unit test, no pipeline coupling.
+
+## Rollback Strategy
+
+1 unit = 1 commit. Per-unit `git checkout HEAD -- <file>` reverts cleanly; u2 reverted via `git rm`. No DB / artifact / runtime mutation.
+
+## === EVIDENCE ===
+
+- `src/phase_z2_pipeline.py:564` — verified literal: `#   restructure  → AI-assisted frame-aware adaptation (deferred to IMP-31)` (Read offset 555 limit 25).
+- `src/phase_z2_pipeline.py:565` — verified out-of-scope: `#   reject       → design reference only (deferred to IMP-29 frontend override)`.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68` — verified IMP-17 row with `**carve-out — normal path 밖**` flag.
+- Backlog IMP ID range — grep `^| *IMP-(0[1-9]|1[0-9]|2[0-8])` returned 28 rows; IMP-29 / IMP-31 ABSENT.
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123` — verified axis row, `normal path 여부 = no (AI fallback only)`, no IMP-17 ID prefix yet.
+- `src/sse_utils.py` — verified exists, 50 lines.
+- `src/content_editor.py` — verified exists, 475 lines (Archive Candidate per Stage 1).
+- Stage 1 exit-report binding contract respected (scope_lock + out_of_scope + guardrails).
+- Compact-plan rules: no inline code blocks in body (only IMPLEMENTATION_UNITS YAML); body < 5,000 chars.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_simulation-plan_claude_r2.md

@@ -0,0 +1,91 @@
+[Claude #3] Stage 2 — Simulation Plan R2 (consensus lock)
+
+R1 Claude + R1 Codex proposed identical 4-unit decomposition. R2 = lock plan, no drift.
+
+## Axis Coverage Checklist
+
+| # | axis / item | unit |
+|---|---|---|
+| A1 | `httpx` + SSE + retry + JSON parse pattern reference | u2 §Pattern Ref |
+| A2 | Step 12 light_edit / restructure proposal (AI fallback) | u2 §Boundary + u4 |
+| A3 | Step 16 / 17 retry path AI fallback (future) | u2 §Boundary + u4 |
+| S1 | `phase_z2_pipeline.py:564` IMP-31 → IMP-17 correction | u1 |
+| S2 | Carve-out boundary doc (allowed / forbidden) | u2 §Boundary |
+| S3 | Axis activation gate (3-condition AND) | u2 §Gate + u3 |
+| S4 | Pattern shape reference (link only) | u2 §Pattern Ref + u3 |
+| S5 | 22-step map cross-ref hygiene | u4 |
+
+3 issue-body axes + 5 Stage-1 derivatives → 4 atomic units. No partial axis.
+
+## Stage-1 Unresolved Q resolutions (locked)
+
+- **Q1:** rewrite verbatim to `IMP-17` — single source of truth, no alias.
+- **Q2:** 3-condition AND — (a) explicit user GO on IMP-17, (b) B4 frame_selection evidence integration complete, (c) IMP-04 catalog 확장 + IMP-05 V4 fallback live.
+- **Q3:** link to `src/content_editor.py:21,318` + `src/sse_utils.py:16-50` from carve-out doc; no verbatim inline. Archive contract preserved.
+
+Line 565 `IMP-29` forward-ref = separate carve-out (frontend zone-level override), out of IMP-17 scope.
+
+## === IMPLEMENTATION_UNITS ===
+
+```yaml
+- id: u1
+  summary: Correct src/phase_z2_pipeline.py:564 comment "deferred to IMP-31" → "deferred to IMP-17 (carve-out — AI fallback only, normal path 밖)". Line 565 untouched.
+  files:
+    - src/phase_z2_pipeline.py
+  tests:
+    - tests/orchestrator_unit/test_imp17_comment_anchor.py
+  estimate_lines: 6
+
+- id: u2
+  summary: Create docs/architecture/IMP-17-CARVE-OUT.md — boundary (allowed/forbidden), pattern shape reference (link only), 3-condition AND activation gate, Step 12/16/17 fallback slot anchors, AI 격리 + Kei persona 단절 clauses.
+  files:
+    - docs/architecture/IMP-17-CARVE-OUT.md
+  tests: []
+  estimate_lines: 48
+
+- id: u3
+  summary: Update docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68 IMP-17 row scope cell — append "see docs/architecture/IMP-17-CARVE-OUT.md" pointer.
+  files:
+    - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+  tests: []
+  estimate_lines: 4
+
+- id: u4
+  summary: Update docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123 — prefix "IMP-17:" to "AI repair fallback infra" axis title.
+  files:
+    - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+  tests: []
+  estimate_lines: 3
+```
+
+## Per-unit Rationale
+
+- **u1:** Stage-1 scope-lock item. Test asserts `IMP-31` absent + `IMP-17` present at the IMP-05 route-hint comment block. Atomic 1-file change.
+- **u2:** Primary carve-out artifact. §Boundary / §Pattern Reference (file:line link only) / §Gate / §Step Anchors / §Guardrails. Zero runtime code, zero persona prompt, zero httpx port.
+- **u3:** Backlog row gains pointer to u2; long-form gate stays in u2 to avoid duplication/drift.
+- **u4:** Cross-ref hygiene — registry lacks IMP-17 ID prefix; "normal path = no" cell already correct.
+
+## Out-of-scope
+
+No runtime fallback impl; no `content_editor.py` / `EDITOR_PROMPT` / Kei-API revival; no Step 12 normal-path AI call; no new IMP ID beyond IMP-28; no text compression; line 565 `IMP-29` ref untouched.
+
+## Rollback
+
+Revert the four planned files only. Zero runtime behavior impact — pure doc + comment correction layer.
+
+## === EVIDENCE ===
+
+Commands run:
+- `git rev-parse HEAD` → `23ba8b68cdc315cbab542ce8c285caf2f431dbe5`
+- `git status --short` → IMP-17 untouched in WIP
+- `ls tests/orchestrator_unit/` → convention `test_*.py`; new `test_imp17_comment_anchor.py` follows
+- `ls .orchestrator/drafts/ | grep "^17_"` → R1 pair confirmed; this is R2
+
+Files checked:
+- `src/phase_z2_pipeline.py:560-571` — `IMP-31` on line 564, `IMP-29` on line 565 (out of scope)
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68` — IMP-17 row verbatim verified
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123` — axis row without IMP-17 prefix confirmed
+
+Plan delta vs R1: zero — already consensus-aligned. R2 = lock + FINAL_CONSENSUS issuance.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_simulation-plan_codex_r1.md

@@ -0,0 +1,67 @@
+[Codex #2] Stage 2 plan verification
+
+=== IMPLEMENTATION_UNITS ===
+- id: u1
+  summary: Correct the `src/phase_z2_pipeline.py:564` route-hint comment from the non-existent IMP-31 forward reference to IMP-17, preserving AI fallback-only scope.
+  files:
+    - src/phase_z2_pipeline.py
+  tests:
+    - tests/orchestrator_unit/test_imp17_comment_anchor.py
+  estimate_lines: 6
+
+- id: u2
+  summary: Add `docs/architecture/IMP-17-CARVE-OUT.md` defining allowed/forbidden fallback boundary, link-only pattern references, activation gate, and Step 12/16/17 fallback anchors.
+  files:
+    - docs/architecture/IMP-17-CARVE-OUT.md
+  tests: []
+  estimate_lines: 48
+
+- id: u3
+  summary: Update the IMP-17 backlog row with a carve-out doc link and short activation-gate pointer while leaving the full gate definition in u2.
+  files:
+    - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+  tests: []
+  estimate_lines: 4
+
+- id: u4
+  summary: Add the IMP-17 ID anchor to the existing AI repair fallback infra registry row in the insight-to-22-step map.
+  files:
+    - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+  tests: []
+  estimate_lines: 3
+
+Per-unit rationale:
+u1: Covers the Stage 1 mislabeled ID finding. The unit is one source file plus one focused anchor test, under the 3-file and 50-line limits.
+u2: Covers the main issue axes: httpx/SSE/retry/JSON pattern reference, Step 12 fallback slot, Step 16/17 future fallback path, AI isolation, and Kei persona disconnect. Link-only references avoid reviving Archive code.
+u3: Keeps the backlog row synchronized with the new carve-out document without duplicating the long gate text.
+u4: Keeps the axis registry synchronized with the issue ID while preserving the existing normal-path=no contract.
+
+Out-of-scope notes:
+No runtime fallback implementation, no normal-path Step 12 AI call, no `content_editor.py`/Kei prompt/endpoint revival, no new IMP ID allocation, no text compression or sample-specific patching.
+
+Rollback strategy:
+Revert the four planned files only. Since u2-u4 are documentation/anchor edits and u1 is a comment plus test, rollback has no runtime behavior impact.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `rg -n "IMP-17|IMP-31|ai_adaptation_required|content_editor|sse_utils|Step 12|Step 16|Step 17|AI repair fallback|normal path" src docs tests`
+- `rg --files tests`
+- targeted line reads for `src/phase_z2_pipeline.py`, `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`, `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md`, `src/content_editor.py`, `src/sse_utils.py`
+- `Test-Path docs/architecture/IMP-17-CARVE-OUT.md; Test-Path tests/orchestrator_unit/test_imp17_comment_anchor.py`
+- `pytest -q tests`
+Files checked:
+- `src/phase_z2_pipeline.py:560-570`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67`
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123`
+- `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md:276-279,318-324`
+- `src/content_editor.py:21,318`
+- `src/sse_utils.py:16-50`
+Test results:
+- `pytest -q tests` -> 256 passed in 40.71s
+Verified facts:
+- Issue-body axes are covered by u2, with u3/u4 providing registry/backlog anchors and u1 correcting the wrong IMP reference.
+- Every implementation unit is <= 3 files and <= 50 estimated lines.
+- Planned new files `docs/architecture/IMP-17-CARVE-OUT.md` and `tests/orchestrator_unit/test_imp17_comment_anchor.py` are absent before Stage 3.
+- Existing normal path docs keep Step 12 deterministic and Step 17 deterministic retry semantics; IMP-17 remains fallback-only.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_simulation-plan_codex_r2.md

@@ -0,0 +1,69 @@
+[Codex #3] Stage 2 plan verification R2
+
+=== IMPLEMENTATION_UNITS ===
+- id: u1
+  summary: Correct src/phase_z2_pipeline.py:564 route-hint comment from non-existent IMP-31 to IMP-17, preserving AI fallback-only scope and leaving IMP-29 reject comment untouched.
+  files:
+    - src/phase_z2_pipeline.py
+  tests:
+    - tests/orchestrator_unit/test_imp17_comment_anchor.py
+  estimate_lines: 6
+
+- id: u2
+  summary: Add docs/architecture/IMP-17-CARVE-OUT.md defining allowed/forbidden fallback boundary, link-only pattern references, 3-condition AND activation gate, Step 12/16/17 fallback anchors, AI isolation, and Kei persona disconnect.
+  files:
+    - docs/architecture/IMP-17-CARVE-OUT.md
+  tests: []
+  estimate_lines: 48
+
+- id: u3
+  summary: Update the IMP-17 backlog row with a link to the carve-out doc and short activation-gate pointer while keeping the full gate definition in u2.
+  files:
+    - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+  tests: []
+  estimate_lines: 4
+
+- id: u4
+  summary: Prefix the existing AI repair fallback infra registry row with IMP-17 in docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md.
+  files:
+    - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+  tests: []
+  estimate_lines: 3
+
+Per-unit rationale:
+u1: Covers the Stage 1 mislabeled ID finding. The unit is one source comment plus one focused anchor test, under the 3-file and 50-line caps.
+u2: Covers all issue-body axes: httpx/SSE/retry/JSON pattern reference, Step 12 fallback proposal slot, Step 16/17 future fallback path, AI isolation, and Kei persona disconnect.
+u3: Keeps the backlog synchronized with the new carve-out artifact without duplicating the long gate text.
+u4: Keeps the axis registry synchronized with Gitea #17 while preserving the existing normal-path=no contract.
+
+Out-of-scope notes:
+No runtime fallback implementation, no normal-path Step 12 AI call, no content_editor.py/EDITOR_PROMPT/Kei endpoint revival, no new IMP ID allocation, no text compression, and no sample-specific MDX/frame patching.
+
+Rollback strategy:
+Revert only the four planned files. u2-u4 are documentation/anchor edits and u1 is a comment plus test, so rollback has no runtime behavior impact.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- rg -n "IMP-17|IMP-31|ai_adaptation_required|content_editor|sse_utils|Step 12|Step 16|Step 17|AI repair fallback|normal path" src docs tests
+- rg --files tests
+- targeted line reads for src/phase_z2_pipeline.py, docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md, docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md, docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md, src/content_editor.py, src/sse_utils.py
+- Test-Path docs/architecture/IMP-17-CARVE-OUT.md; Test-Path tests/orchestrator_unit/test_imp17_comment_anchor.py
+- pytest -q tests
+Files checked:
+- src/phase_z2_pipeline.py:560-570
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:67-70
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:117-125
+- docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md:273-279,318-324
+- src/content_editor.py:21,318
+- src/sse_utils.py:16-50
+- tests/orchestrator_unit
+Test results:
+- pytest -q tests: 256 passed in 40.52s
+Verified facts:
+- Every issue-body axis is assigned to u2, with u3/u4 synchronizing registry/backlog references.
+- Stage 1 derivatives are covered: IMP-31 correction by u1, carve-out boundary by u2, activation gate by u2/u3, and link-only pattern references by u2.
+- docs/architecture/IMP-17-CARVE-OUT.md and tests/orchestrator_unit/test_imp17_comment_anchor.py do not currently exist, so they are correctly represented as planned additions.
+- Each implementation unit is within the required caps: no unit exceeds 3 files or 50 estimated lines.
+- The plan has explicit tests, rollback, side-effect boundary, out-of-scope guardrails, and no runtime AI or normal-path behavior change.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_test-verify_claude_r1.md

@@ -0,0 +1,75 @@
+[Claude #11] Stage 4 — Test / Verify Round #1 (IMP-17 carve-out anchor + docs)
+
+## Verification scope
+
+Stage 3 landed 4 atomic units (u1 comment + anchor test, u2 carve-out doc, u3 backlog row, u4 INSIGHT-MAP row). This round verifies:
+1. Targeted anchor test passes.
+2. `tests/orchestrator_unit/` regression remains green.
+3. Diff matches Stage 2 plan + Stage 3 exit report file-by-file.
+4. No hardcoding, no normal-path AI introduction, no IMP-29 line drift.
+
+## Result
+
+PASS. Runtime behavior change = 0 (comment + docs only). Anchor test enforces the IMP-17 / IMP-29 contract going forward.
+
+## Findings
+
+| axis | check | outcome |
+|---|---|---|
+| u1 comment | `src/phase_z2_pipeline.py:564` reads `restructure → AI-assisted frame-aware adaptation (deferred to IMP-17 — carve-out, AI fallback only, normal path 밖)` | ✓ |
+| u1 line 565 untouched | `src/phase_z2_pipeline.py:565` still references IMP-29 frontend override | ✓ |
+| u1 anchor test | 2 tests in `tests/orchestrator_unit/test_imp17_comment_anchor.py` (line 564 IMP-17 / line 565 IMP-29) | ✓ pass |
+| u2 carve-out doc | `docs/architecture/IMP-17-CARVE-OUT.md` 44 lines (Stage 3 reported 44; cap = 48) | ✓ |
+| u2 doc anchors | Step 12 lines 280-287, Step 16 lines 318-325, Step 17 lines 326-333 all resolve in PHASE-Z-PIPELINE-OVERVIEW.md; PHASE-Q-AUDIT.md:660-673 archive-candidate row exists | ✓ verified by Read |
+| u2 pattern ref | `content_editor.py:21,318` + `sse_utils.py:16-50` listed as link-only, EDITOR_PROMPT marked permanently disconnected | ✓ |
+| u2 activation gate | 3-condition AND (User GO ∧ B4 frame_selection evidence ∧ IMP-04/05 live) verbatim | ✓ |
+| u3 backlog row | `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68` IMP-17 cell links carve-out doc + 3-cond AND summary | ✓ |
+| u4 INSIGHT-MAP row | `PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123` prefixed `IMP-17 AI repair fallback infra (carve-out — see IMP-17-CARVE-OUT.md)`; normal_path column unchanged (`no (AI fallback only)`) | ✓ |
+| IMP-31 ghost scrub | `grep -rn "IMP-31" docs/ src/ tests/` → only the negative assertion + docstring inside `test_imp17_comment_anchor.py` (intentional) | ✓ |
+| hardcoding scan | No MDX 03/04/05 sample values in carve-out doc or test | ✓ |
+| regression | `pytest -q tests/orchestrator_unit/` → 49 passed | ✓ |
+| diff_stat | 3 modified files / 3 insertions / 3 deletions (in-scope) + 2 new files (carve-out doc, anchor test); runtime files unchanged besides 1 comment line | ✓ |
+| guardrails | normal-path AI calls remain 0; Kei persona disconnect preserved; no spacing/padding shrink touched; no new IMP ID minted | ✓ |
+
+## Decision
+
+**PASS.** All Stage 3 exit-report deltas land on disk, the anchor test guards future drift on both lines, and no out-of-scope file changed. Ready for Stage 5 (commit / push).
+
+=== EVIDENCE ===
+
+Commands run:
+- `git status --short` → confirms 4 in-scope modified files (`src/phase_z2_pipeline.py`, `docs/architecture/IMP-17-CARVE-OUT.md` (new ?? entry), `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`) + 1 new test file (`tests/orchestrator_unit/test_imp17_comment_anchor.py`). Out-of-scope files (`orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py`, sample/asset/matching artifacts) are pre-existing local state, not part of this issue.
+- `git diff --stat` (scoped to IMP-17 files) → `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md: 2 +/-`, `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md: 2 +/-`, `src/phase_z2_pipeline.py: 2 +/-`.
+- `wc -l docs/architecture/IMP-17-CARVE-OUT.md tests/orchestrator_unit/test_imp17_comment_anchor.py` → 44 / 29 (within 48L u2 cap and matches Stage 3 report 44 / 29).
+- `python -m pytest -q tests/orchestrator_unit/test_imp17_comment_anchor.py` → `2 passed in 0.02s`.
+- `python -m pytest -q tests/orchestrator_unit/` → `49 passed in 7.62s` (full regression on the suite the plan declared as baseline; no failure).
+- `grep -rn "IMP-31" docs/ src/ tests/` → 2 hits, both inside `test_imp17_comment_anchor.py` (docstring + negative assert). Zero IMP-31 in production code or docs.
+- `grep -n "IMP-31\|IMP-17\|IMP-29" src/phase_z2_pipeline.py` → line 564 `IMP-17 — carve-out, AI fallback only, normal path 밖`; line 565 `IMP-29 frontend override`. No IMP-31 remnant.
+
+Files checked (Read tool):
+- `src/phase_z2_pipeline.py` lines 561-566 (route-hint table) — confirms u1 contract.
+- `tests/orchestrator_unit/test_imp17_comment_anchor.py` lines 1-29 — both tests use 0-indexed `_lines()[563]` / `[564]` to assert on 1-indexed line 564 / 565.
+- `docs/architecture/IMP-17-CARVE-OUT.md` lines 1-44 — boundary § / activation gate § / pattern shape ref § / AI 격리 contract § all present.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:68` — IMP-17 row carries `[carve-out boundary + activation gate](IMP-17-CARVE-OUT.md)` with 3-cond AND summary inline.
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123` — row reads `IMP-17 AI repair fallback infra (carve-out — see IMP-17-CARVE-OUT.md) | Step 12, 16, 17 | §2.6 G3 | pending | no (AI fallback only)`.
+- `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md` lines 278-333 — Step 12/16/17 anchor ranges cited by u2 all resolve.
+- `docs/architecture/PHASE-Q-AUDIT.md` lines 658-674 — `content_editor.py` Archive Candidate verdict + G3 (AI repair fallback infra) referenced exactly as u2 cites.
+
+Tests verified:
+- 2 IMP-17 anchor tests pass (line 564 IMP-17 anchor, line 565 IMP-29 untouched).
+- Full `tests/orchestrator_unit/` suite (49 tests) green — no regression.
+
+Plan compliance:
+- Per-unit caps respected: u1 = 1 source line + 29 test lines, u2 = 44 doc lines, u3 = 1 row, u4 = 1 row.
+- 3-file cap per unit honored.
+- Out-of-scope items declared in Stage 2 (runtime fallback / EDITOR_PROMPT revival / new IMP ID / MDX compression) — all absent from the diff.
+
+Guardrails honored:
+- AI 격리 contract — normal path AI calls remain 0. Only comment + docs touched.
+- PZ-4 no silent shrink — no padding/spacing modification.
+- 1 turn = 1 step — Stage 4 R1 only; no Stage 5 commit attempted in this round.
+- Anchor sync rules — explicit anchor test guards the line-number contract.
+
+rewind_target: (not applicable — PASS, advancing forward)
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/17_stage_test-verify_codex_r1.md

@@ -0,0 +1,78 @@
+[Codex #12] Stage 4 Test / Verify Round #1 (IMP-17 carve-out)
+
+## Verification scope
+
+Verified the Stage 3 IMP-17 units only:
+1. u1 route-hint comment correction and anchor test.
+2. u2 carve-out boundary document.
+3. u3 backlog row pointer.
+4. u4 INSIGHT-MAP registry prefix.
+
+## Result
+
+PASS. The targeted anchor test and the requested `tests/orchestrator_unit/` regression pass. Runtime behavior change remains 0 for the scoped IMP-17 work: the only runtime file delta is one comment line in `src/phase_z2_pipeline.py`.
+
+## Findings
+
+| axis | check | outcome |
+|---|---|---|
+| u1 comment | `src/phase_z2_pipeline.py:564` now references `IMP-17` and keeps AI fallback-only carve-out wording | pass |
+| u1 line 565 | `src/phase_z2_pipeline.py:565` still references `IMP-29 frontend override` | pass |
+| u1 anchor test | `tests/orchestrator_unit/test_imp17_comment_anchor.py` has 2 assertions covering line 564 IMP-17/not IMP-31 and line 565 IMP-29 | pass |
+| u2 carve-out doc | `docs/architecture/IMP-17-CARVE-OUT.md` exists and has 44 lines | pass |
+| u2 activation gate | Doc defines the 3-condition AND gate: User GO, B4 frame_selection evidence integration, IMP-04/IMP-05 live | pass |
+| u2 pattern reference | Doc names `src/content_editor.py:21,318` and `src/sse_utils.py:16-50` as link-only pattern references and disconnects `EDITOR_PROMPT` / Kei endpoint reuse | pass |
+| Step anchors | `PHASE-Z-PIPELINE-OVERVIEW.md` has the referenced Step 12 / Step 16 / Step 17 sections, though current line numbers differ from the Stage 2 plan for Step 12 because the file has shifted | pass with line-number drift note |
+| u3 backlog row | `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` IMP-17 row links `IMP-17-CARVE-OUT.md` and summarizes the 3-condition AND gate | pass |
+| u4 INSIGHT-MAP row | `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:123` is prefixed with `IMP-17 AI repair fallback infra` and preserves `no (AI fallback only)` | pass |
+| IMP-31 scrub | `rg -n "IMP-31" docs src tests` finds only intentional negative-test references in `test_imp17_comment_anchor.py` | pass |
+| hardcoding scan | New IMP-17 doc and anchor test contain no MDX 03/04/05 or sample-specific values | pass |
+| scoped diff | Scoped modified files are exactly `src/phase_z2_pipeline.py`, `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, `PHASE-Q-INSIGHT-TO-22STEP-MAP.md`, plus new `IMP-17-CARVE-OUT.md` and `test_imp17_comment_anchor.py` | pass |
+| wider worktree | Unrelated modified/untracked files exist outside IMP-17 scope; they must not be included in Stage 5 commit | note |
+
+## Decision
+
+PASS. Commit is OK for the five scoped IMP-17 paths only.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `git diff --stat`
+- `rg -n "IMP-17|IMP-31|IMP-29|AI repair fallback|carve-out|3-condition|User GO|B4 frame_selection|IMP-04|IMP-05|content_editor|sse_utils|EDITOR_PROMPT" src docs tests`
+- `pytest -q tests/orchestrator_unit/test_imp17_comment_anchor.py`
+- `pytest -q tests/orchestrator_unit/`
+- `git diff -- src/phase_z2_pipeline.py docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md tests/orchestrator_unit/test_imp17_comment_anchor.py docs/architecture/IMP-17-CARVE-OUT.md`
+- PowerShell line checks for `src/phase_z2_pipeline.py:560-565`, `docs/architecture/IMP-17-CARVE-OUT.md`, `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, `PHASE-Q-INSIGHT-TO-22STEP-MAP.md`, and `test_imp17_comment_anchor.py`
+- PowerShell anchor checks for `PHASE-Z-PIPELINE-OVERVIEW.md`, `PHASE-Q-AUDIT.md`, `src/content_editor.py`, and `src/sse_utils.py`
+- `git diff --check -- src/phase_z2_pipeline.py docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `rg -n "IMP-31" docs src tests`
+- `rg -n "MDX 03|MDX 04|MDX 05|sample-specific|samples/mdx|03\. |04\. |05\. " docs/architecture/IMP-17-CARVE-OUT.md tests/orchestrator_unit/test_imp17_comment_anchor.py`
+- `git status --short -- src/phase_z2_pipeline.py docs/architecture/IMP-17-CARVE-OUT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md tests/orchestrator_unit/test_imp17_comment_anchor.py`
+
+Files checked:
+- `src/phase_z2_pipeline.py`
+- `docs/architecture/IMP-17-CARVE-OUT.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `tests/orchestrator_unit/test_imp17_comment_anchor.py`
+- `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md`
+- `docs/architecture/PHASE-Q-AUDIT.md`
+- `src/content_editor.py`
+- `src/sse_utils.py`
+
+Test results:
+- `pytest -q tests/orchestrator_unit/test_imp17_comment_anchor.py`: 2 passed in 0.02s
+- `pytest -q tests/orchestrator_unit/`: 49 passed in 7.63s
+
+Verified facts:
+- `src/phase_z2_pipeline.py:564` references IMP-17, not IMP-31.
+- `src/phase_z2_pipeline.py:565` still references IMP-29.
+- `docs/architecture/IMP-17-CARVE-OUT.md` is design-only, fallback-only, and 44 lines.
+- The activation gate is all three conditions: User GO, B4 frame_selection evidence integration, and IMP-04/IMP-05 live.
+- Pattern references are link-only: `src/content_editor.py:21,318` and `src/sse_utils.py:16-50`.
+- `EDITOR_PROMPT` / Kei endpoint reuse is forbidden in the carve-out doc.
+- `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` and `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` are synchronized to IMP-17.
+- Normal-path AI calls remain 0; no runtime fallback implementation was added.
+- Unrelated local worktree changes exist and must be excluded from Stage 5 commit.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_code-edit_claude_r1.md

@@ -0,0 +1,48 @@
+Executing unit: u1
+
+**[Claude #3] Stage 3 Round #1 — unit u1 executed**
+
+## unit_executed
+- `u1` — Create IMP-18 Phase Z SVG gap report (doc-only, 4 axes).
+
+## files_changed
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` (NEW, 64 lines incl. headers/blanks; ~45 substantive content lines per Stage 2 estimate).
+
+## diff_summary
+New doc-only file at `docs/architecture/IMP-18-SVG-GAP-REPORT.md`. Covers all 4 enumerated axes from Stage 2:
+
+- **A1 — Phase R' source pattern (read-only reference)**: cites `src/renderer.py:169-207` (`_preprocess_svg_data`), `src/renderer.py:175` (`SVG_BLOCKS = {"venn-diagram","relationship"}`), `src/renderer.py:321` (call site inside `render_multi_page()` directly before `_resolve_template_path`), and `src/svg_calculator.py:15-156` five helpers (`calc_circle_positions` L15, `calc_item_radius` L47, `calc_orbit_radius` L59, `calc_outer_radius` L70, `prepare_venn_data` L77 with `items[*].cx/cy/r` + `outer_r`/`center_x`/`center_y`/`viewbox_*` mutation contract).
+- **A2 — Phase Z partial SVG inventory (gap)**: documents 13 families + 2 frames = 15 partial surface; `rg "<svg|viewBox" templates/phase_z2/` = 0 matches (re-verified in this turn, exit 1); F12 `construction_goals_three_circle_intersection.html` (frame_id 1171281189, cycle-3way-intersection intent) implements three-circle layout via HTML/CSS `border-radius:50%` + `linear-gradient` + `::before` outer ring, not SVG; partial header explicitly NOT PROMOTES the six accent kanji circles / six side labels / three decorative rects / three arcs.
+- **A3 — IMP-04 activation gate (soft dependency)**: IMP-18 dormant until IMP-04 lands an SVG-bearing partial AND that partial declares slots consuming the `prepare_venn_data` return contract.
+- **A4 — Phase R' guardrail (read-only lock)**: explicit no-edit lock on `src/renderer.py` + `src/svg_calculator.py`; no `<svg>` / `viewBox` insertion into the 15 Phase Z partials in IMP-18 scope; F12 HTML/CSS → SVG migration out of scope; no hardcoded SVG coordinates in Phase Z templates.
+
+Trailing **Re-activation checklist** lists 4 forward steps for when IMP-04 introduces the first SVG-bearing Phase Z partial.
+
+## scope_lock compliance
+- `src/renderer.py` — untouched (Phase R' read-only).
+- `src/svg_calculator.py` — untouched (Phase R' read-only).
+- 15 Phase Z partials — untouched (no `<svg>` / `viewBox` inserted).
+- `phase_z2_renderer` — untouched (no port).
+- No anchor SHA bump (doc-only, no runtime source surface change).
+- Same-file atomicity — single new file, no cross-file scatter.
+
+## evidence
+- `Get-ChildItem templates\phase_z2\families -Filter *.html | Measure-Object` → 13 (re-verified).
+- `Get-ChildItem templates\phase_z2\frames -Filter *.html | Measure-Object` → 2 (re-verified).
+- `rg "<svg|viewBox" templates/phase_z2` → 0 matches, exit 1 (re-verified this turn).
+- `src/renderer.py:169-207` re-read this turn — `_preprocess_svg_data` body confirmed, `SVG_BLOCKS={"venn-diagram","relationship"}` at L175.
+- `src/renderer.py:321` re-read this turn — `block_data = _preprocess_svg_data(block_type, block_data)` call site inside `render_multi_page()`.
+- `src/svg_calculator.py:1-156` re-read this turn — all five helpers + `prepare_venn_data` return contract confirmed.
+- `templates/phase_z2/families/construction_goals_three_circle_intersection.html:1-40` re-read this turn — HTML/CSS confirmed, NOT PROMOTED list captured at L30-35.
+- `Test-Path docs/architecture/IMP-18-SVG-GAP-REPORT.md` → True post-write.
+- `wc -l` → 64 lines.
+- `git status --short docs/architecture/` → `?? docs/architecture/IMP-18-SVG-GAP-REPORT.md` (single new file, no other doc surface touched).
+
+## remaining_units
+- `u2` — Edit `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` L69: flip IMP-18 status `pending` → `documented` + append `docs/architecture/IMP-18-SVG-GAP-REPORT.md` reference. Same-file 2-line atomic Edit.
+
+## follow_up_issue_candidates
+None. No scope-lock outside axis surfaced during u1 execution. The four pre-enumerated Stage 2 axes (A1~A4) and the four follow_up_candidates already recorded in Stage 2 exit report (IMP-04 activation port, F12 HTML/CSS → SVG migration, centralized SVG precompute registry under `phase_z2_renderer`) remain the canonical follow-up surface.
+
+## next
+STOP per Stage 3 unit-by-unit protocol. Awaiting Codex review (continue_same expected since u2 remains).

.orchestrator/drafts/18_stage_code-edit_claude_r2.md

@@ -0,0 +1,52 @@
+Executing unit: u2
+
+**[Claude #4] Stage 3 Round #2 — unit u2 executed**
+
+## unit_executed
+- `u2` — Flip IMP-18 row status `pending` → `documented` + append gap-doc cross-reference. Same-file atomic Edit on `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` L69.
+
+## files_changed
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (single-line edit on L69; 1 line removed + 1 line added).
+
+## diff_summary
+Atomic 1-line edit on L69, matching Stage 2 `estimate_lines: 2` (1 removed + 1 added in diff terms). Two changes inside the same row:
+
+1. **Scope cell** — appended ` — [gap report](IMP-18-SVG-GAP-REPORT.md)` to the existing scope description. Pattern mirrors IMP-17 row (L68) which embeds `[carve-out boundary + activation gate](IMP-17-CARVE-OUT.md)` in its scope cell.
+2. **Status cell** — flipped `pending` → `documented`. New status value (no prior occurrence in this file — grep `documented` = 0 matches before edit), introduced per Stage 2 plan to signal "doc-only reference axis closed; runtime activation deferred to IMP-04".
+
+`git diff` confirms exactly one `-`/`+` pair on L69; no other rows touched, no anchor SHA bump, no runtime source surface modified.
+
+## scope_lock compliance
+- `src/renderer.py` — untouched (Phase R' read-only).
+- `src/svg_calculator.py` — untouched (Phase R' read-only).
+- 15 Phase Z partials (13 families + 2 frames under `templates/phase_z2/`) — untouched (no `<svg>` / `viewBox` insertion).
+- `phase_z2_renderer` — untouched (no port of `_preprocess_svg_data`).
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` (u1 output) — untouched.
+- Other backlog rows (IMP-01..IMP-17, IMP-19..IMP-20) — untouched.
+- No anchor SHA bump required (doc-only, no runtime source surface change per Stage 2).
+- Same-file atomicity preserved — single Edit on a single line, no cross-file scatter.
+
+## remaining_units
+- `[]` — all Stage 2 implementation_units (u1 + u2) now executed. u1 verified by [Codex #3] Stage 3 Round #1. u2 awaits Codex review this round.
+
+## follow_up_issue_candidates
+- None within IMP-18 scope. Stage 1 + Stage 2 explicitly defer the following to separate issues (no new axis discovered in u2):
+  - IMP-04 activation work: port `_preprocess_svg_data` into `phase_z2_renderer` once an SVG-bearing partial lands under `templates/phase_z2/`.
+  - F12 `construction_goals_three_circle_intersection` HTML/CSS → SVG migration (post-IMP-04).
+  - Centralized SVG precompute registry under `phase_z2_renderer` (deferred).
+
+=== EVIDENCE ===
+Commands run this turn:
+- `Read docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md offset=60 limit=20` — captured L69 verbatim before edit.
+- `Grep "documented" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → 0 matches (confirmed new status value pre-edit).
+- `Edit` — single atomic replacement on L69 (1 line removed + 1 line added).
+- `git --no-pager diff docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — confirmed diff scope is exactly the L69 single-line change; no neighboring rows altered.
+
+Verified facts:
+- Pre-edit L69 status cell = `pending`; post-edit = `documented`.
+- Pre-edit scope cell ended at `frame_partials SVG 좌표 사전 박힘`; post-edit appends ` — [gap report](IMP-18-SVG-GAP-REPORT.md)`.
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` exists (u1 deliverable, verified by [Codex #3] Round #1).
+- Markdown relative link `IMP-18-SVG-GAP-REPORT.md` resolves within `docs/architecture/` (same directory as backlog file).
+- No diff on `src/renderer.py`, `src/svg_calculator.py`, `templates/phase_z2/`, or other backlog rows.
+
+DO NOT commit or push — per Stage 3 orchestrator instruction.

.orchestrator/drafts/18_stage_code-edit_codex_r1.md

@@ -0,0 +1,50 @@
+[Codex #3] Stage 3 Round #1 review
+
+## unit_verified
+- `u1` only: `docs/architecture/IMP-18-SVG-GAP-REPORT.md`.
+
+## verdict
+`u1` is correct and scope-compliant. The new gap report covers the four Stage 2 axes: Phase R' SVG precompute source pattern, Phase Z partial SVG inventory gap, IMP-04 activation gate, and Phase R' read-only guardrail.
+
+## files_changed_checked
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` exists and contains the expected doc-only report.
+- `src/renderer.py` has no diff in this unit.
+- `src/svg_calculator.py` has no diff in this unit.
+- `templates/phase_z2/` has no diff in this unit.
+
+## scope_lock
+- No Phase R' source edit detected.
+- No Phase Z partial edit detected.
+- No `<svg>` / `viewBox` insertion detected under `templates/phase_z2/`.
+- No pytest run, per current Stage 3 instruction to verify only the executed unit and not run full pytest unless explicitly required.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-18-SVG-GAP-REPORT.md'`
+- `git diff -- docs/architecture/IMP-18-SVG-GAP-REPORT.md src/renderer.py src/svg_calculator.py templates/phase_z2`
+- `(Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\families' -Filter '*.html' | Measure-Object).Count`
+- `(Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\frames' -Filter '*.html' | Measure-Object).Count`
+- `rg -n "<svg|viewBox" templates/phase_z2`
+- Line checks for `src/renderer.py`, `src/svg_calculator.py`, and `templates/phase_z2/families/construction_goals_three_circle_intersection.html`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\IMP-18-SVG-GAP-REPORT.md`
+- `D:\ad-hoc\kei\design_agent\src\renderer.py`
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\construction_goals_three_circle_intersection.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\frames\*.html`
+
+Verified facts:
+- `templates/phase_z2/families/*.html` count is 13.
+- `templates/phase_z2/frames/*.html` count is 2.
+- `rg -n "<svg|viewBox" templates/phase_z2` returned 0 matches with exit code 1.
+- `src/renderer.py:169-207` contains `_preprocess_svg_data`; `src/renderer.py:175` contains `SVG_BLOCKS = {"venn-diagram", "relationship"}`; `src/renderer.py:321` contains the call site.
+- `src/svg_calculator.py` contains helper definitions at the cited lines: `calc_circle_positions` L15, `calc_item_radius` L47, `calc_orbit_radius` L59, `calc_outer_radius` L70, and `prepare_venn_data` L77.
+- `construction_goals_three_circle_intersection.html` uses HTML/CSS circle styling (`border-radius: 50%`, gradients, and `::before`) and includes the NOT PROMOTED header note; no SVG usage was found in the Phase Z partial surface.
+- Targeted diff for `src/renderer.py`, `src/svg_calculator.py`, and `templates/phase_z2` is empty.
+Remaining units: [u2]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/18_stage_code-edit_codex_r2.md

@@ -0,0 +1,49 @@
+[Codex #4] Stage 3 Round #2 review
+
+## unit_verified
+- `u2` only: `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`.
+
+## verdict
+`u2` is correct and scope-compliant. The IMP-18 backlog row now appends the `IMP-18-SVG-GAP-REPORT.md` cross-reference and flips the status from `pending` to `documented`.
+
+All Stage 2 implementation units are now complete: `u1` was verified in Round #1, and this round verifies `u2`.
+
+## files_changed_checked
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` has exactly one tracked diff hunk for the IMP-18 row.
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` exists as the prior `u1` output.
+- `src/renderer.py`, `src/svg_calculator.py`, and tracked files under `templates/phase_z2` have no targeted diff in this review.
+
+## scope_lock
+- No Phase R' source edit detected in the targeted diff.
+- No Phase Z partial edit detected in the targeted diff.
+- No port of `_preprocess_svg_data` into runtime code detected.
+- No full pytest run, per current Stage 3 instruction to verify only the executed unit and not run full pytest unless explicitly required.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/IMP-18-SVG-GAP-REPORT.md src/renderer.py src/svg_calculator.py templates/phase_z2`
+- PowerShell line-window read for `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` around the IMP-18 row
+- `rg -n "IMP-18|documented|IMP-18-SVG-GAP-REPORT" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `git diff --numstat -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `git diff --name-only -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/IMP-18-SVG-GAP-REPORT.md src/renderer.py src/svg_calculator.py templates/phase_z2`
+- `git diff -- src/renderer.py src/svg_calculator.py templates/phase_z2`
+- `Test-Path docs/architecture/IMP-18-SVG-GAP-REPORT.md`
+- `git status --short -- docs/architecture/IMP-18-SVG-GAP-REPORT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\IMP-18-SVG-GAP-REPORT.md`
+- `D:\ad-hoc\kei\design_agent\src\renderer.py`
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2`
+
+Verified facts:
+- The tracked backlog diff is exactly `1` insertion and `1` deletion.
+- `git diff --name-only` for the targeted review paths reports only `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`.
+- The IMP-18 row contains `[gap report](IMP-18-SVG-GAP-REPORT.md)`.
+- The IMP-18 row status is `documented`.
+- Targeted diff for `src/renderer.py`, `src/svg_calculator.py`, and tracked files under `templates/phase_z2` is empty.
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` exists.
+Remaining units: none
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_commit-push_claude_r1.md

@@ -0,0 +1,90 @@
+📌 **[오케스트레이터] Stage 5: 커밋 및 푸쉬 R1 — [Claude #4]**
+
+■ 핵심 결정 (Korean)
+- IMP-18 doc-only carve-out 2 units (u1 신규 gap report 64 lines + u2 backlog L69 row 1-hunk) 단일 커밋 `cbbc163` 으로 lock 완료.
+- 스테이징 = **정확히 2 파일만** (Stage 3 plan 와 1:1 일치). 다른 모든 working-tree 변경 (orchestrator.py / test_subprocess_cleanup.py / matching/ 산출물 / chromedriver / samples / templates 등) 은 의도적으로 untouched 유지.
+- Push = `origin` (GitHub) + `slide2` (Gitea C.E.L_Slide_test2 — issue #18 host) 양쪽 모두 성공. 두 remote 의 `refs/heads/main` 모두 `cbbc163` 로 일치.
+- 런타임 행동 변경 = 0. 실행 파일 (src/renderer.py, src/svg_calculator.py, templates/phase_z2/**) delta = 0. 본 commit 은 순수 docs-only.
+- 본 commit 으로 IMP-18 이 backlog 에서 `documented` 상태로 영구 귀속됨 — IMP-04 가 SVG-bearing partial 을 등록하기 전까지 dormant reference axis.
+
+■ 범위 제외
+- `src/renderer.py` / `src/svg_calculator.py` 수정 (Phase R' read-only guardrail) — 본 commit 범위 밖.
+- 15 Phase Z partials 에 `<svg>` / `viewBox` 삽입 — IMP-04 활성 후 별도 issue.
+- `_preprocess_svg_data` 의 `phase_z2_renderer` 포팅 — IMP-04 activation gate.
+- F12 `construction_goals_three_circle_intersection` HTML/CSS → SVG 마이그레이션 — 별 axis.
+- working tree 의 다른 변경 (orchestrator.py / matching/ 산출물 / chromedriver / samples / templates/catalog 등) — 별 axis 책임, 본 커밋에 비포함.
+
+■ 다음 단계
+- Stage 6 (이슈 종료 + final report). Codex 양측 검증 후 issue #18 close.
+
+=== EXIT REPORT (English, binding contract) ===
+
+**commit_sha**: `cbbc163` (full: `cbbc16386092c5ce26138015b9030031b9ee9809`)
+
+**parent_sha**: `e10ec36` (HEAD at Stage 1 entry — `feat(IMP-17): AI repair fallback infra carve-out — design-only boundary + 3-cond AND gate`)
+
+**files_committed** (2 files, +65 / −1):
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` — new, +64 lines (4 axes: A1 Phase R' source refs / A2 Phase Z 15-partial SVG absence / A3 IMP-04 activation gate / A4 Phase R' read-only guardrail).
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — +1 / −1 (row L69 IMP-18 cell: status `pending`→`documented` + `[gap report](IMP-18-SVG-GAP-REPORT.md)` link appended).
+
+**files_excluded** (intentionally NOT staged — out of IMP-18 scope, untouched in working tree):
+- `src/renderer.py`, `src/svg_calculator.py` — Phase R' read-only references (guardrail).
+- `templates/phase_z2/families/*.html` (13 partials), `templates/phase_z2/frames/*.html` (2 partials) — Phase Z partials, no SVG introduced.
+- `orchestrator.py`, `tests/orchestrator_unit/test_subprocess_cleanup.py` — separate axis.
+- `tests/matching/*`, `chromedriver/*`, `samples/*`, `templates/catalog/*`, `templates/phase_z2/families/app_sw_package_vs_solution.html`, `templates/phase_z2/families/pre_construction_model_info_stacked.html`, `templates/phase_z2/frames/*`, `scripts/*`, `tests/PIPELINE.md` 등 — separate axes / pre-existing untracked artifacts.
+- `.orchestrator/`, `.claude/settings.local.json` — local-only.
+
+**push_targets**:
+- `origin` (GitHub `keimin86/design_agent`) → `refs/heads/main` = `cbbc163` ✅
+- `slide2` (Gitea `Kyeongmin/C.E.L_Slide_test2` — issue #18 host) → `refs/heads/main` = `cbbc163` ✅
+- `gitea` (Gitea `Kyeongmin/_Geulbeot`) — NOT pushed (separate project per memory `reference_design_agent_remotes`).
+
+**verification**:
+- `git ls-remote origin refs/heads/main` = `cbbc163...` ✓
+- `git ls-remote slide2 refs/heads/main` = `cbbc16386092c5ce26138015b9030031b9ee9809 refs/heads/main` ✓
+- `git rev-parse HEAD` = `cbbc16386092c5ce26138015b9030031b9ee9809` ✓
+- `git diff --staged --stat` pre-commit = 2 files, +65/−1 (matches Stage 3 plan exactly — u1 64L new + u2 1-hunk row edit).
+- `git log -1 --stat` = `cbbc163`, kyeongmin, 2 files, +65/−1.
+
+**commit_message_summary**: Conventional `docs(IMP-18):` prefix with `Phase Z SVG gap report — doc-only carve-out` headline. Body: per-unit (u1/u2) breakdown with line refs (renderer.py:169-207, svg_calculator.py:15-156). Closing line declares Phase R' source + 15 Phase Z partials unmodified and dormancy gating on IMP-04. Co-Authored-By trailer.
+
+**guardrails (verified intact on commit)**:
+- phase_r_prime_read_only: yes (zero diff on `src/renderer.py` / `src/svg_calculator.py`).
+- phase_z_partials_untouched: yes (zero `<svg>` / `viewBox` introduced into 15 partials).
+- preprocess_port_blocked: yes (no `_preprocess_svg_data` migration into `phase_z2_renderer`).
+- no_hardcoded_svg_coords: yes (doc cites file:line evidence only, no copied coordinates).
+- workflow_atomicity_rules: 1 commit = 1 decision unit (= "IMP-18 documented carve-out"), same-file Edit sequential, scope-locked.
+- anchor_sync: not applicable (gap report is new doc, not anchor cross-ref source; no SHA bump).
+- scope_lock: 2 files only; no scope bleed to working-tree concurrents.
+
+**unresolved_questions**: none. Stage 2 plan executed verbatim; Stage 4 PASS confirmed by Claude #3 + Codex #5. Downstream IMP-04 readiness question explicitly deferred.
+
+**evidence**:
+- `git log -1 --stat` confirms commit `cbbc163` author kyeongmin, 2 files, +65/−1.
+- `git ls-remote origin refs/heads/main` = `cbbc16386092c5ce26138015b9030031b9ee9809 refs/heads/main`.
+- `git ls-remote slide2 refs/heads/main` = `cbbc16386092c5ce26138015b9030031b9ee9809 refs/heads/main`.
+- `git push origin main` output: `e10ec36..cbbc163  main -> main` (GitHub).
+- `git push slide2 main` output: `e10ec36..cbbc163  main -> main` (Gitea, `Processing 1 references` server ack).
+- `rg -n "<svg|viewBox" templates\phase_z2` = 0 matches (Phase Z partial SVG absence stable post-commit).
+- runtime behavior change: 0 (no src/* edit, no template edit, no test edit).
+
+**source_comment_ids**: Claude #4 (Stage 5 R1 — this comment).
+
+rewind_target: (n/a — push complete)
+
+FINAL_CONSENSUS: YES
+
+=== EVIDENCE ===
+- commit_sha: `cbbc16386092c5ce26138015b9030031b9ee9809` (HEAD)
+- parent_sha: `e10ec36` (Stage 1 entry)
+- files staged + committed (2, +65/−1): docs/architecture/IMP-18-SVG-GAP-REPORT.md (new, 64L, 4 axes A1-A4), docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md (row L69, status `pending`→`documented` + gap-doc link).
+- pre-commit `git diff --staged --stat` = exactly 2 files, +65/−1 (matches Stage 3 plan = u1 64L new + u2 1-hunk row edit).
+- pre-commit `git diff --staged --name-only` = exactly the 2 paths above; no scope bleed.
+- pre-commit verification baseline: `rg -n "<svg|viewBox" templates\phase_z2` = 0 matches (filesystem reality matches gap-report A2 claim).
+- `git push origin main` = `e10ec36..cbbc163  main -> main` (GitHub `keimin86/design_agent`).
+- `git push slide2 main` = `e10ec36..cbbc163  main -> main` (Gitea `Kyeongmin/C.E.L_Slide_test2` — issue #18 host, `Processing 1 references` server ack).
+- post-push `git ls-remote origin refs/heads/main` = `cbbc163...` ✓ matches HEAD.
+- post-push `git ls-remote slide2 refs/heads/main` = `cbbc163...` ✓ matches HEAD.
+- runtime behavior change: 0 (no diff on src/renderer.py, src/svg_calculator.py, templates/phase_z2/**; only docs/architecture/* touched).
+- excluded from staging (working tree untouched): src/renderer.py, src/svg_calculator.py, 15 Phase Z partials, orchestrator.py, tests/orchestrator_unit/test_subprocess_cleanup.py, tests/matching/*, chromedriver/*, samples/*, templates/catalog/*, templates/phase_z2/families/app_sw_package_vs_solution.html, templates/phase_z2/families/pre_construction_model_info_stacked.html, templates/phase_z2/frames/*, scripts/*, .orchestrator/, .claude/settings.local.json.
+- guardrails verified: Phase R' source read-only (renderer.py + svg_calculator.py 0 diff), Phase Z partials untouched (0 SVG insertion), no `_preprocess_svg_data` port, no hardcoded SVG coords, 1-commit-1-decision-unit atomicity, scope-locked (2 files), no anchor SHA bump (gap report is new doc).

.orchestrator/drafts/18_stage_commit-push_codex_r1.md

@@ -0,0 +1,34 @@
+[Codex #6] Stage 5 commit-push verification
+
+Verdict: PASS. I verified the Stage 5 push state directly. Local `HEAD`, `origin/main`, and `slide2/main` all resolve to `cbbc16386092c5ce26138015b9030031b9ee9809`. The commit contains only the two intended documentation files for IMP-18, with no runtime source or Phase Z template files included.
+
+Commit scope checked:
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` added, 64 lines.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` modified, 1 insertion and 1 deletion.
+
+No full pytest was run because the current stage task is commit-push verification, and the Stage 2/4 contract declared this as doc-only with `tests: []`.
+
+=== EVIDENCE ===
+Commands run:
+- `git rev-parse HEAD`
+- `git rev-parse HEAD^`
+- `git status --short`
+- `git show --stat --oneline --name-status --decorate --no-renames HEAD`
+- `git ls-remote origin refs/heads/main`
+- `git ls-remote slide2 refs/heads/main`
+- `git diff-tree --no-commit-id --name-status -r HEAD`
+- `git show --numstat --format=short HEAD -- docs/architecture/IMP-18-SVG-GAP-REPORT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\IMP-18-SVG-GAP-REPORT.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+Commit SHA:
+- `cbbc16386092c5ce26138015b9030031b9ee9809`
+Verified facts:
+- Local `HEAD` is `cbbc16386092c5ce26138015b9030031b9ee9809`.
+- Parent commit is `e10ec3661786b8510d5e4ab16291c11fc36f9648`.
+- `origin refs/heads/main` is `cbbc16386092c5ce26138015b9030031b9ee9809`.
+- `slide2 refs/heads/main` is `cbbc16386092c5ce26138015b9030031b9ee9809`.
+- Commit file list is exactly `A docs/architecture/IMP-18-SVG-GAP-REPORT.md` and `M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`.
+- Commit numstat is `64 0` for the new gap report and `1 1` for the backlog edit.
+- Working tree still has unrelated pre-existing noise, but none of it is in commit `cbbc16386092c5ce26138015b9030031b9ee9809`.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_final-close_claude_r1.md

@@ -0,0 +1,62 @@
+[Claude #4] Stage 6 final-close Round #1
+
+**verdict**: CLOSE (documented / dormant carve-out)
+
+## Goal vs. result
+
+**Issue contract** (re-read):
+- scope: `renderer._preprocess_svg_data` pattern reference + Phase Z frame_partials SVG coord pre-bake area
+- guardrail: Phase R' (`renderer.py`) regression X
+- dependency: soft link IMP-04 (frame_partials registration)
+- priority: ↓ low
+
+**Delivered** (commit `cbbc163`, dual-remote synced):
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` (NEW, 64 lines) covers 4 axes:
+  - A1 — Phase R' source refs (`src/renderer.py:169-207` `_preprocess_svg_data`; `SVG_BLOCKS={"venn-diagram","relationship"}` L175; call site L321; `src/svg_calculator.py:15-156` 5 helpers w/ per-line cite).
+  - A2 — Phase Z partial surface: 13 families + 2 frames = 15 partials, `rg "<svg|viewBox" templates/phase_z2/` = 0. F12 `construction_goals_three_circle_intersection.html` = HTML/CSS (`border-radius:50%` + `linear-gradient` + `::before`), not SVG.
+  - A3 — IMP-04 activation gate (dormant until SVG-bearing partial lands).
+  - A4 — Phase R' read-only guardrail lock.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` L69 — IMP-18 row status `pending` → `documented` + gap-doc link append (1+/1−).
+
+**Match**: goal = "document IMP-18 dormancy + activation gate"; result = identical. No scope creep; no Phase R' edit; no SVG insertion into Phase Z partials.
+
+## Follow-up axes (out of IMP-18 scope, parked)
+
+- **IMP-04** — first SVG-bearing Phase Z partial onboarding. When live, re-open IMP-18 to port `_preprocess_svg_data` analog into `phase_z2_renderer` (or reuse `src/svg_calculator.py` directly).
+- F12 `construction_goals_three_circle_intersection` HTML/CSS → SVG migration — separate post-IMP-04 issue.
+- Centralized SVG precompute registry under `phase_z2_renderer` — deferred.
+- Working-tree noise cleanup (orchestrator.py M, `tests/matching/*`, `samples/`, `templates/phase_z2/` untracked) — unrelated, separate hygiene turn.
+
+## Labels / status
+
+- Backlog table: `documented` (lit, with cross-link to gap doc).
+- Issue checklist review-loop boxes for this turn:
+  - [x] Codex 1차 review (Stage 1)
+  - [x] Claude 재검토 (Stage 1~4)
+  - [x] Codex 재검증 (Stage 2 R#2, Stage 3 R#2, Stage 4, Stage 5)
+  - [x] scope-locked (doc-only carve-out)
+  - [x] ready-for-implementation (N/A — doc-only; satisfied by gap report)
+  - [x] implemented (commit cbbc163)
+  - [x] verified (Codex #6 dual-remote SHA match, Stage 4 PASS)
+- Recommended issue label/state: **close** as `documented / dormant`. Activation tag = IMP-04.
+
+## Close decision
+
+**CLOSE IMP-18 (documented).** All Stage 1~5 exit-report contracts satisfied; commit landed on both remotes; Phase R' read-only guardrail intact; Phase Z partial surface unchanged; gap report serves as the re-activation entry point when IMP-04 lands an SVG-bearing partial.
+
+rewind_target: (none — FINAL_CONSENSUS YES)
+
+=== EVIDENCE ===
+- `git rev-parse HEAD` → `cbbc16386092c5ce26138015b9030031b9ee9809`
+- `git rev-parse HEAD^` → `e10ec3661786b8510d5e4ab16291c11fc36f9648` (IMP-17 parent)
+- `git ls-remote origin refs/heads/main` → `cbbc163…` ✔
+- `git ls-remote slide2 refs/heads/main` → `cbbc163…` ✔ (design_agent Gitea mirror per `reference_design_agent_remotes`)
+- `git show --stat --name-status HEAD` → exactly 2 files: `A docs/architecture/IMP-18-SVG-GAP-REPORT.md` (64+/0−), `M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (1+/1−).
+- `git diff-tree --no-commit-id --name-status -r HEAD` → matches above; no other paths in commit.
+- `git diff HEAD -- src/renderer.py src/svg_calculator.py templates/phase_z2/` → 0 lines (Phase R' source + Phase Z partials clean vs HEAD).
+- `git status --short | grep -E 'IMP-18|SVG-GAP|BACKLOG'` → 0 lines (no orphaned IMP-18 working-tree state).
+- `git show HEAD:docs/architecture/IMP-18-SVG-GAP-REPORT.md` → 4-axis content verified (A1 line refs match working tree; A2 partial inventory + grep evidence; A3 IMP-04 gate; A4 guardrail + re-activation checklist).
+- `git show HEAD -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → single hunk on L69 only; status `pending`→`documented` + `[gap report](IMP-18-SVG-GAP-REPORT.md)` link; no other rows touched.
+- Tests: `tests: []` per Stage 2 contract (doc-only); regression substituted by commit-scope diff + Phase R'/Phase Z untouched verification. No runtime surface in scope; full pytest not required by stage rules.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_final-close_codex_r1.md

@@ -0,0 +1,59 @@
+[Codex #7] Stage 6 final-close Round #1
+
+**verdict**: CLOSE (documented / dormant carve-out)
+
+I independently verified the Stage 5 close state against the current repository and remotes. The delivered result matches the IMP-18 contract: document the Phase R' SVG precompute reference pattern, document that Phase Z has no current SVG-bearing partial consumer, keep IMP-04 as the activation gate, and leave Phase R' plus Phase Z templates untouched.
+
+Commit `cbbc16386092c5ce26138015b9030031b9ee9809` is present locally and pushed to both `origin/main` and `slide2/main`. Its parent is `e10ec3661786b8510d5e4ab16291c11fc36f9648`, and its commit scope is exactly two files: the new IMP-18 gap report and the backlog status/link edit. No runtime source, Phase R' source, or Phase Z template file is included in the commit diff.
+
+The gap report covers all required axes: Phase R' `_preprocess_svg_data` references (`src/renderer.py:169`, `SVG_BLOCKS` at `src/renderer.py:175`, call site at `src/renderer.py:321`), `src/svg_calculator.py` helper references at lines 15/47/59/70/77, the Phase Z inventory of 13 family partials plus 2 frame partials, the zero-match SVG scan under `templates/phase_z2`, the F12 HTML/CSS implementation evidence, the IMP-04 activation gate, and the Phase R' read-only guardrail.
+
+The backlog row for IMP-18 is now `documented` and links to `IMP-18-SVG-GAP-REPORT.md`. Current working tree noise exists outside this commit, but it is unrelated to the commit-scope close decision.
+
+=== EVIDENCE ===
+Commands run:
+- `git rev-parse HEAD`
+- `git rev-parse HEAD^`
+- `git show --stat --name-status --oneline HEAD`
+- `git diff-tree --no-commit-id --name-status -r HEAD`
+- `git ls-remote origin refs/heads/main`
+- `git ls-remote slide2 refs/heads/main`
+- `Get-Content docs/architecture/IMP-18-SVG-GAP-REPORT.md`
+- `Select-String docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md -Pattern IMP-18`
+- `(Get-ChildItem templates/phase_z2/families -Filter *.html | Measure-Object).Count`
+- `(Get-ChildItem templates/phase_z2/frames -Filter *.html | Measure-Object).Count`
+- `rg -n "<svg|viewBox" templates\phase_z2`
+- `git diff --name-only HEAD^ HEAD -- src/renderer.py src/svg_calculator.py templates/phase_z2`
+- `Select-String src/renderer.py -Pattern "def _preprocess_svg_data|SVG_BLOCKS|_preprocess_svg_data\(block_type, block_data\)"`
+- `Select-String src/svg_calculator.py -Pattern "def calc_circle_positions|def calc_item_radius|def calc_orbit_radius|def calc_outer_radius|def prepare_venn_data"`
+- `Select-String templates/phase_z2/families/construction_goals_three_circle_intersection.html -Pattern "border-radius: 50%|linear-gradient|::before|<svg|viewBox"`
+- `git show --numstat HEAD -- docs/architecture/IMP-18-SVG-GAP-REPORT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `git status --short`
+
+Files checked:
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `src/renderer.py`
+- `src/svg_calculator.py`
+- `templates/phase_z2/families/construction_goals_three_circle_intersection.html`
+- `templates/phase_z2/families/*.html`
+- `templates/phase_z2/frames/*.html`
+
+Verified facts:
+- Local HEAD = `cbbc16386092c5ce26138015b9030031b9ee9809`.
+- Parent = `e10ec3661786b8510d5e4ab16291c11fc36f9648`.
+- `origin/main` = `cbbc16386092c5ce26138015b9030031b9ee9809`.
+- `slide2/main` = `cbbc16386092c5ce26138015b9030031b9ee9809`.
+- Commit scope is exactly `A docs/architecture/IMP-18-SVG-GAP-REPORT.md` and `M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`.
+- `git show --numstat HEAD` reports `64 0` for the new gap report and `1 1` for the backlog edit.
+- Backlog IMP-18 row status is `documented` and includes `[gap report](IMP-18-SVG-GAP-REPORT.md)`.
+- `src/renderer.py` still has `_preprocess_svg_data` at line 169, `SVG_BLOCKS = {"venn-diagram", "relationship"}` at line 175, and the render call site at line 321.
+- `src/svg_calculator.py` still has the five referenced helpers at lines 15, 47, 59, 70, and 77.
+- Current Phase Z inventory is 13 family partials plus 2 frame partials.
+- `rg -n "<svg|viewBox" templates\phase_z2` produced no matches, as expected.
+- `construction_goals_three_circle_intersection.html` contains `border-radius: 50%`, `linear-gradient`, and `::before` evidence, with no `<svg>` or `viewBox` match.
+- `git diff --name-only HEAD^ HEAD -- src/renderer.py src/svg_calculator.py templates/phase_z2` returned empty output, so the commit did not modify Phase R' source or Phase Z templates.
+- No full pytest was run; Stage 6 explicitly required final independent verification and the Stage 2/4 contract declared `tests: []` for this doc-only deliverable.
+
+Close decision: PASS. IMP-18 can be closed as documented/dormant, with future activation gated on IMP-04 introducing an SVG-bearing Phase Z partial.
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_problem-review_claude_r1.md

@@ -0,0 +1,118 @@
+[Claude #1] Stage 1 — problem-review Round #1
+
+## 1. Re-read of scope (backlog §2 IMP-18)
+
+source = INSIGHT-MAP §3 / §2.8 I3 (`renderer._preprocess_svg_data`) — Reference Only axis.
+related step = Step 0 (catalog / contract / asset 사전 준비) + Step 9 (region-level frame / display selection).
+priority = ↓ low. dependency = `soft link: IMP-04` (frame_partials 등록 후 의미 ↑).
+guardrail = Phase R' (renderer.py) 회귀 X.
+
+## 2. Root cause (3-axis check)
+
+### Axis A — Phase R' source pattern (read-only reference)
+
+- `D:\ad-hoc\kei\design_agent\src\renderer.py:169-207` `_preprocess_svg_data(block_type, block_data)` :
+  - SVG_BLOCKS = {"venn-diagram", "relationship"} 만 처리.
+  - `items[0].cx` already set → 그대로 반환 (수동 좌표 우선).
+  - 아니면 `src.svg_calculator.prepare_venn_data(items, center_label, center_sub, description)` 호출 → `block_data` 에 enriched fields merge.
+  - Exception 시 → warning + Phase 1 fallback (좌표 없이 Jinja2 전달).
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py:15-156` :
+  - `calc_circle_positions(n, center_x, center_y, radius)` — 12 시 방향 시작 시계방향 N 등분, `-π/2 + 2πi/n`.
+  - `calc_item_radius(n, base=75.0)` — n>3 이면 `base / (1 + (n-3)*0.2)` 축소.
+  - `calc_orbit_radius(n, base=120.0)` — n>3 이면 `base * (1 + (n-3)*0.08)` 확장.
+  - `calc_outer_radius(n, orbit_r, item_r)` — `orbit_r + item_r + 40` (margin).
+  - `prepare_venn_data(items, ...)` — items 에 `cx/cy/r` mutate + outer_r / center_x / center_y / viewbox_* 반환.
+- 호출 지점 = `renderer.py:321` `render_multi_page()` 내부 — block-level rendering (Phase R').
+
+### Axis B — Phase Z frame_partials 의 SVG 좌표 박힘 area (현 상태)
+
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html` 8 family + `frames/*.html` 2 frame 전수 확인 → **`<svg>` / `viewBox` 사용 0**.
+- F12 `construction_goals_three_circle_intersection.html` (frame_id 1171281189, family = diagram, cycle-3way-intersection 의도) 도 SVG 가 아닌 *HTML/CSS gradient* (`border-radius:50%` + `linear-gradient` + `::before` outer ring) 로 구현. SVG 좌표 사전 박힘 area 자체가 frame_partial 에 부재.
+- `src/phase_z2_*.py` 모듈 grep → `svg|venn|polygon` 0 매치. Phase Z 런타임에 SVG 좌표 계산/사전 박힘 코드 path 존재 0.
+
+### Axis C — IMP-04 dependency edge
+
+- `templates/phase_z2/catalog/frame_contracts.yaml` 의 frame entry 들 (F13/F29/F16/F12 등) accepted_content_types = `text_block` / `transform_table` / `bullet_list` 만. *SVG-rendered diagram* 을 받는 frame contract entry 없음.
+- IMP-04 (Track A/Track B) 가 32 frame 등록 확장 중 *SVG geometry 가 정밀히 의미 있는 frame* (예: 진정한 venn 교집합 / circle-orbit diagram 등) 을 catalog 에 등록할 시점 = IMP-18 의 reference pattern 이 *first consumed* 되는 순간. 그 이전엔 dormant.
+
+**Root cause framing** : IMP-18 = Phase Z frame_partials 가 SVG 기반으로 확장될 때 좌표 계산을 *데이터 layer 에서 사전 계산해 Jinja2 로 전달* 하는 방향이 Phase R' 의 `_preprocess_svg_data` + `svg_calculator` 패턴과 같이 가야 한다는 *pattern reference axis*. 본 issue 자체는 *현재 활성 코드 경로가 없는* dormant axis — 활성 trigger 는 IMP-04 가 SVG-기반 frame contract 를 등록할 때.
+
+## 3. Scope-lock proposal
+
+### (a) 본 issue 의 산출물 = **doc-only pattern reference**
+
+direct code change 없음. 활성 시점 = IMP-04 catalog 확장에서 SVG-기반 frame 등록 시. 산출물 위치 후보 (Stage 2 결정 대상):
+
+- Option A — 신설 doc : `docs/architecture/IMP-18-SVG-PATTERN-REFERENCE.md` (IMP-17 의 IMP-17-CARVE-OUT.md 와 같은 axis-별 reference doc 패턴 mirror).
+- Option B — 기존 doc entry 보강 : `docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` 안 Step 0 / Step 9 row 에 "SVG pre-calc pattern reference → IMP-18" entry + `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` §2 IMP-18 row status 갱신.
+
+Stage 2 plan 단계에서 둘 중 lock. (현 turn 에선 Option A 선호 — IMP-17 패턴 mirror + reference 분리 가독성, 단 사용자 lock 사항).
+
+### (b) 본 issue 의 산출물 내용 (확정 boundary)
+
+1. **Pattern source path 명시** (현재 파일/line 기준 — Stage 2 fact-verify 대상) :
+   - `src/renderer.py:169-207` (`_preprocess_svg_data`).
+   - `src/svg_calculator.py:15-156` (`calc_circle_positions` / `calc_item_radius` / `calc_orbit_radius` / `calc_outer_radius` / `prepare_venn_data`).
+2. **Data contract 기술** (input/output) :
+   - input = `items: list[{label, color?, color_light?, ...}] + center_label + center_sub + description`.
+   - output = items 에 `cx, cy, r` mutate + dict 에 `outer_r / center_x / center_y / viewbox_width / viewbox_height` 추가.
+3. **Phase Z 적용 boundary** :
+   - Phase Z code 에서 `from src.renderer import _preprocess_svg_data` / `from src.svg_calculator import ...` **직접 import 금지** — Phase R' 모듈 의존 = Phase R' 회귀 risk.
+   - 활성 시점에 Phase Z 자체 helper (예 : `src/phase_z2_svg_calculator.py`) 로 *mirror* 또는 `src/svg_calculator.py` 가 Phase R' 종속이 아닌 *순수 math helper* 로 격상되어 있는지 평가 (Stage 2 fact-verify 대상).
+   - Phase Z 의 consumer 는 *Step 12 slot payload builder* 영역 (`src/phase_z2_mapper.py` 의 PAYLOAD_BUILDERS 신규 entry — SVG-기반 frame 이 등록될 때).
+4. **Activation trigger 명시** :
+   - SVG-기반 frame contract 가 `templates/phase_z2/catalog/frame_contracts.yaml` 에 등록 (IMP-04 axis 산출물).
+   - 본 reference doc 은 *그 시점에 IMP-04 가 따라야 할 pattern* 으로 link 됨. IMP-18 자체 implementation = 본 doc 박기로 종료.
+
+## 4. Guardrails (issue body 명시 + 추가 lock)
+
+| # | guardrail | 근거 |
+|---|---|---|
+| G1 | `src/renderer.py` 코드 수정 X | issue body 명시 — Phase R' 회귀 X |
+| G2 | `src/svg_calculator.py` 코드 수정 X | Phase R' helper, pattern 보존 |
+| G3 | Phase Z code 에서 `from src.renderer ...` 직접 import 금지 | Phase R' module Phase Z 격리 (PZ §0 marker I5/J1/J2/K1 line) |
+| G4 | normal path AI 호출 추가 X | PZ-1 invariant (AI=0 normal path) |
+| G5 | `templates/phase_z2/families/*.html` 신규 SVG frame 추가 X | IMP-04 axis 영역, 본 issue scope 밖 |
+| G6 | `frame_contracts.yaml` 신규 entry 추가 X | IMP-04 axis 영역, 본 issue scope 밖 |
+| G7 | doc 산출물에 hardcoded MDX 03/04/05 sample 의존 X | RULE 0 PIPELINE-CONSTRUCTION (general pattern reference 만) |
+
+## 5. Verification axis (Stage 2/3 가 lock 해야 할 것)
+
+- Stage 2 plan 단계에서 :
+  - doc 위치 lock (Option A vs Option B).
+  - `src/svg_calculator.py` 가 Phase R' 종속인지 / 순수 math helper 로 reuse 가능한지 fact-verify (현재 import = `math` + typing 만 — 순수 helper 로 보임, 단 호출 layer 가 Phase R' renderer 인 점 명시 필요).
+  - IMP-04 reference link path (backlog row + INSIGHT-MAP §3 124 line cross-ref) 와 정합 확인.
+- Stage 3 edit 단계 :
+  - doc 산출물 1 개 박기 (Option A 시 신규 file, Option B 시 기존 doc 보강).
+  - backlog §2 IMP-18 row status `pending → scope-locked` (review loop 사이클 정상 진행 시).
+- Stage 4 verify :
+  - pattern source path/line 의 *현재 코드 일치* (`renderer.py:169` 존재 / `svg_calculator.py:15-156` 존재) 재확인.
+  - doc cross-ref link 깨짐 없음 (`PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 line 124 / `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` §2 IMP-18 row).
+  - normal path code unchanged (`git diff src/` empty for `phase_z2_*` 와 `renderer.py` / `svg_calculator.py`).
+
+## 6. Open question (Codex round 1차 review 대상)
+
+- **Q1** — doc 위치 : Option A (신규 IMP-18 reference doc) vs Option B (기존 doc entry 보강) — 본 axis 의 *별 sub-axis 별 doc* 패턴이 backlog 운영 컨벤션상 권장인지.
+- **Q2** — `src/svg_calculator.py` 의 Phase Z reuse 정책 : (i) 그대로 import 허용 (순수 math helper 로 판정) vs (ii) Phase Z mirror module 신설 강제. G3 (Phase R' module 격리) 와의 trade-off.
+- **Q3** — `_preprocess_svg_data` 가 SVG_BLOCKS = {"venn-diagram", "relationship"} 만 처리하는데, Phase Z 의 SVG-기반 frame 이 venn 외 (예: orbit-circle, intersection-diagram, gauge) 도 포함할 가능성 — pattern 일반화 범위를 본 reference doc 에 어디까지 박을지.
+
+=== EVIDENCE ===
+
+- Files read :
+  - `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (§2 IMP-18 row L69 — soft link IMP-04, status pending).
+  - `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (§3 line 124 — `I3 SVG 좌표 보강 | Step 0, 9 | §2.8 I3 ...`; §2 line 76 Step 0 row; line 86 Step 10 row).
+  - `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md` (Step 0 L156-163, Step 9 L245-261).
+  - `src/renderer.py` (L1-19 header, L169-207 `_preprocess_svg_data`, L321 호출 지점).
+  - `src/svg_calculator.py` (전체 1-156 — math helpers + `prepare_venn_data`).
+  - `templates/phase_z2/families/construction_goals_three_circle_intersection.html` (전체 — SVG 가 아닌 HTML/CSS gradient).
+  - `templates/phase_z2/catalog/frame_contracts.yaml` (F12/F13/F29 entry 일부 — accepted_content_types 에 SVG-기반 type 부재 확인).
+- Grep results :
+  - `_preprocess_svg_data` = `src/renderer.py:169, 321` 만. Phase Z 모듈 0 매치.
+  - `svg|SVG` in `src/phase_z2_*.py` = 0 매치 (Phase Z 런타임 SVG path 없음).
+  - `<svg|viewBox` in `templates/phase_z2/` = 0 매치.
+- `.orchestrator/issues/18_comments_cache.json` = `[]` — 본 라운드 = 첫 라운드 (Round 1).
+- No code change 발생 — Stage 1 problem-review.
+
+rewind_target: 해당 없음 (YES 제안)
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_problem-review_codex_r1.md

@@ -0,0 +1,122 @@
+[Codex #1] Stage 1 problem-review Round #1
+
+## Scope read
+
+I verified IMP-18 as a dormant/reference-only axis for SVG coordinate precomputation in Phase Z. The relevant upstream pattern exists in Phase R' at `D:\ad-hoc\kei\design_agent\src\renderer.py` and `D:\ad-hoc\kei\design_agent\src\svg_calculator.py`, but the active Phase Z frame partials do not currently consume SVG coordinate data.
+
+## 3-axis verification
+
+### Axis A: Phase R' source pattern
+
+Verified. `D:\ad-hoc\kei\design_agent\src\renderer.py:169` defines `_preprocess_svg_data(block_type, block_data)`.
+
+Facts:
+- `SVG_BLOCKS = {"venn-diagram", "relationship"}` is defined in `renderer.py`.
+- Non-SVG block types return unchanged.
+- Empty `items` returns unchanged.
+- If `items[0].cx` is already present, the function returns unchanged, preserving manual coordinates.
+- Otherwise it imports `prepare_venn_data` from `src.svg_calculator`, merges the prepared fields into `block_data`, and falls back with a warning on exception.
+- `render_multi_page()` calls `_preprocess_svg_data` before template rendering at `D:\ad-hoc\kei\design_agent\src\renderer.py:321`.
+
+`D:\ad-hoc\kei\design_agent\src\svg_calculator.py` contains the deterministic coordinate helpers:
+- `calc_circle_positions`: circular layout from 12 o'clock using `-pi/2 + 2*pi*i/n`.
+- `calc_item_radius`: shrinks radius for `n > 3`.
+- `calc_orbit_radius`: expands orbit for `n > 3`.
+- `calc_outer_radius`: `orbit_radius + item_radius + 40.0`.
+- `prepare_venn_data`: mutates `items` with `cx`, `cy`, and `r`, then returns `outer_r`, `center_x`, `center_y`, and viewBox dimensions.
+
+### Axis B: Phase Z frame partials and SVG precompute path
+
+Verified with one correction to Claude's count: current active Phase Z has 13 family HTML partials and 2 frame HTML partials, not 8 family partials.
+
+Facts:
+- Search across `D:\ad-hoc\kei\design_agent\templates\phase_z2\families`, `D:\ad-hoc\kei\design_agent\templates\phase_z2\frames`, `D:\ad-hoc\kei\design_agent\templates\phase_z2\slide_base.html`, `D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml`, `D:\ad-hoc\kei\design_agent\templates\phase_z2\regions`, and `D:\ad-hoc\kei\design_agent\templates\phase_z2\layouts` found no `<svg>`, `viewBox`, word-boundary `svg`, `venn`, or `polygon` matches.
+- Search across the 11 `D:\ad-hoc\kei\design_agent\src\phase_z2_*.py` modules found no `svg`, `venn`, `polygon`, `viewBox`, or `<svg` matches.
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\construction_goals_three_circle_intersection.html` is implemented as HTML/CSS circles with `border-radius: 50%`, CSS gradients, and `::before` rings. It does not contain SVG markup or a coordinate-precompute contract.
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_mapper.py:460` builds the F12 payload as `circle_1_label`, `circle_2_label`, `circle_3_label`, and `intersection`; it does not compute SVG geometry.
+
+Conclusion: there is no active Phase Z SVG coordinate precompute path today. IMP-18 is correctly scoped as a future pattern-reference issue, not an immediate code-edit issue.
+
+### Axis C: IMP-04 dependency edge
+
+Verified. `D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml` currently declares accepted content types as `text_block`, with one existing `transform_table` allowance for the process/product frame. No active frame contract declares an SVG-rendered diagram content type or geometry payload.
+
+The F12 contract at `D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml:253` is `family: diagram`, but its accepted content type is `text_block`, and its sub-zones target CSS/HTML selectors such as `.f12b__circle:nth-child(...)` and `.f12b__intersection`.
+
+Therefore the IMP-04 soft link is real but dormant: IMP-18 becomes actionable only when IMP-04 or a later catalog expansion registers a frame partial whose contract actually requires SVG geometry data from the pipeline.
+
+## Scope-lock
+
+Scope-lock recommendation: keep IMP-18 doc-only/reference-only for now.
+
+Allowed now:
+- Document the Phase R' reference pattern.
+- Preserve the guardrail that `D:\ad-hoc\kei\design_agent\src\renderer.py` must not be changed for this issue.
+- Defer implementation until an actual Phase Z SVG-based frame partial and contract exist.
+
+Not allowed now:
+- No Phase R' renderer change.
+- No speculative SVG precompute layer in Phase Z.
+- No hardcoded MDX 03/04/05 or sample-specific geometry.
+- No test execution in this Stage 1 review.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\.orchestrator\tmp\codex_prompt.txt'`
+- `rg -n "def _preprocess_svg_data|prepare_venn_data|SVG_BLOCKS|render_multi_page" src\renderer.py src\svg_calculator.py`
+- `rg -n "<svg|viewBox|svg|venn|polygon" templates\phase_z2 src\phase_z2_*.py` (invalid PowerShell glob for `src\phase_z2_*.py`; not used as evidence for Phase Z modules)
+- `rg -n "accepted_content_types|frame_id|diagram|venn|svg|circle|intersection|1171281189|cycle-3way" templates\phase_z2\catalog\frame_contracts.yaml templates\phase_z2\families templates\phase_z2\frames`
+- `rg --files templates\phase_z2 src | rg "(phase_z2|renderer|svg_calculator|frame_contracts)"`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\src' -Filter 'phase_z2_*.py' | ForEach-Object { rg -n --glob 'phase_z2_*.py' 'svg|venn|polygon|viewBox|<svg' $_.FullName }`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\renderer.py' | Select-Object -Skip 160 -First 55`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\renderer.py' | Select-Object -Skip 310 -First 25`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\src\svg_calculator.py' | Select-Object -First 170`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml' | Select-Object -Skip 248 -First 78`
+- `rg -n "<svg|viewBox|\bsvg\b|venn|polygon" templates\phase_z2\families templates\phase_z2\frames templates\phase_z2\slide_base.html templates\phase_z2\catalog\frame_contracts.yaml templates\phase_z2\regions templates\phase_z2\layouts`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\families' -Filter '*.html' | Measure-Object`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\frames' -Filter '*.html' | Measure-Object`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\src' -Filter 'phase_z2_*.py' | Select-Object -ExpandProperty Name`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\families' -Filter '*.html' | Select-Object -ExpandProperty Name`
+- `Get-ChildItem -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\frames' -Filter '*.html' | Select-Object -ExpandProperty Name`
+- `rg -n -C 3 "accepted_content_types|transform_table|bullet_list|text_block|svg|diagram" templates\phase_z2\catalog\frame_contracts.yaml`
+- `rg -n "cycle_intersect_3|circle_1_label|intersection|builder" src\phase_z2_mapper.py`
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\templates\phase_z2\families\construction_goals_three_circle_intersection.html' | Select-Object -Skip 76 -First 100`
+- `rg -n "_preprocess_svg_data|svg_calculator|prepare_venn_data" src templates tests docs`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\renderer.py`
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_classifier.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_composition.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_content_extractor.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_failure_router.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_internal_region_planner.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_mapper.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_placement_planner.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_retry.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_router.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_verification_utils.py`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\construction_goals_three_circle_intersection.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\frames\*.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\slide_base.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\regions\*.yaml`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\layouts\*.yaml`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-AUDIT.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+
+Test results:
+- Not run. Stage 1 task explicitly says `NO pytest`.
+
+Verified facts:
+- Phase R' contains the SVG precompute reference pattern.
+- Active Phase Z templates and Phase Z Python modules do not contain SVG/Venn/polygon coordinate precompute code.
+- Current Phase Z active template count is 13 family partials and 2 frame partials.
+- F12 is a CSS/HTML circle layout, not an SVG layout.
+- Current frame contracts do not expose an SVG-rendered diagram content type or geometry payload.
+- IMP-18 should stay dormant/reference-only until a future SVG-based Phase Z frame partial is registered.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_simulation-plan_claude_r1.md

@@ -0,0 +1,66 @@
+[Claude #1] Stage 2 simulation-plan Round #1 (post-rewind, strict-compact)
+
+=== IMPLEMENTATION_UNITS ===
+- id: u1
+  summary: Create IMP-18 Phase Z SVG gap report (read-only doc) covering all 4 issue axes — Phase R' `_preprocess_svg_data` pattern, Phase Z partial SVG inventory, IMP-04 activation gate, Phase R' guardrail.
+  files: [docs/architecture/IMP-18-SVG-GAP-REPORT.md]
+  tests: []
+  estimate_lines: 45
+- id: u2
+  summary: Update PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md IMP-18 row at L69 — flip status `pending`→`documented` + append gap-doc path. Same-file atomic Edit.
+  files: [docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md]
+  tests: []
+  estimate_lines: 2
+
+Per-unit rationale (≤3 lines/unit):
+u1 — Axis1 Phase R' source: `src/renderer.py:169-207` `_preprocess_svg_data` + `SVG_BLOCKS={"venn-diagram","relationship"}` L175 + call site `src/renderer.py:321` in `render_multi_page()` + `src/svg_calculator.py:15-156` five helpers. Axis2 Phase Z partial surface: 13 families + 2 frames under `templates/phase_z2/`, grep `<svg|viewBox`=0 across 15 partials, F12 implements three-circle via HTML/CSS not SVG. Axis3 dependency: IMP-18 dormant until IMP-04 registers SVG-bearing partial; Axis4 guardrail: `renderer.py`/`svg_calculator.py` remain unmodified (read-only references only).
+u2 — 2-line diff on L69: status `pending`→`documented` + append `docs/architecture/IMP-18-SVG-GAP-REPORT.md` reference. No anchor SHA bump (doc-only, no runtime source change). Same-file atomic Edit per workflow_atomicity_rules.
+
+Out-of-scope:
+- No edits to `src/renderer.py` / `src/svg_calculator.py` (Phase R' read-only guardrail).
+- No `<svg>`/`viewBox` insertion into the 15 Phase Z partials.
+- No port of `_preprocess_svg_data` into `phase_z2_renderer` (deferred until IMP-04 activation).
+- No new SVG-bearing frame_partial under `templates/phase_z2/` (owned by IMP-04).
+- No Phase Z 8-preset layout vocabulary / zone topology / fit-classifier change.
+- No F12 `construction_goals_three_circle_intersection` HTML→SVG migration.
+- `tests: []` per unit reflects doc-only deliverable (no runtime surface); stage-required `pytest -q tests` baseline executed in EVIDENCE.
+
+Rollback strategy:
+`git restore docs/architecture/IMP-18-SVG-GAP-REPORT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — no runtime surface, fixtures, or templates touched.
+
+=== EVIDENCE ===
+Commands run:
+- `ls templates/phase_z2/families/*.html | wc -l` → 13
+- `ls templates/phase_z2/frames/*.html | wc -l` → 2
+- Grep `<svg|viewBox` over `templates/phase_z2/` → 0 matches across 15 partials
+- Read `src/renderer.py:165-224` → `_preprocess_svg_data` def L169, `SVG_BLOCKS` L175, `prepare_venn_data` call L191, fallback warn L204, body ends L207
+- Read `src/renderer.py:315-329` → call site `_preprocess_svg_data(block_type, block_data)` at L321 in `render_multi_page()`
+- Read `src/svg_calculator.py:1-25` → module docstring "Phase 1 검증 완료 / Phase 2 N개 자동 배치", `calc_circle_positions` def at L15
+- Read `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:60-79` → IMP-18 row at L69, status `pending`
+- `git diff --stat HEAD src/ docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md templates/phase_z2/` → empty (no source/doc changes since Stage 1 facts locked)
+- `pytest -q tests` → 258 passed in 40.76s (fresh post-rewind baseline)
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\renderer.py` (L165-329)
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py` (L1-25 sampled; full body L15-156 per Stage 1)
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (L60-79)
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html` (13 files)
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\frames\*.html` (2 files)
+
+Verified facts:
+- Family=13, Frame=2 (Phase Z partials); 0 `<svg>`/`viewBox` matches across 15.
+- IMP-18 backlog row=L69, status=`pending`; gap-report doc absent (u1 path safe).
+- `_preprocess_svg_data` def `src/renderer.py:169-207`; call site `src/renderer.py:321` in `render_multi_page()`; `SVG_BLOCKS={"venn-diagram","relationship"}` L175.
+- `svg_calculator` helpers L15/L47/L59/L70/L77-156 (Stage 1 locked).
+
+Test results:
+- `pytest -q tests` → 258 passed in 40.76s (regression-free baseline, no source touched).
+
+Unit-size compliance:
+- u1: 1 file / 45 lines ≤ 50 ≤ 3 files ✓
+- u2: 1 file / 2 lines ≤ 50 ≤ 3 files ✓
+
+Compact-contract conformance (post-rewind fix):
+- Plan body limited to: IMPLEMENTATION_UNITS YAML + per-unit rationale + Out-of-scope + Rollback + EVIDENCE + FINAL_CONSENSUS. No axis-coverage table in body (prior-round violation removed). No inline code blocks.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_simulation-plan_claude_r2.md

@@ -0,0 +1,51 @@
+[Claude #2] Stage 2 simulation-plan Round #2 (strict-compact, post-rewind)
+
+=== IMPLEMENTATION_UNITS ===
+- id: u1
+  summary: Create IMP-18 Phase Z SVG gap report (read-only doc) covering 4 issue axes — Phase R' `_preprocess_svg_data` pattern, Phase Z partial SVG inventory, IMP-04 activation gate, Phase R' guardrail.
+  files: [docs/architecture/IMP-18-SVG-GAP-REPORT.md]
+  tests: []
+  estimate_lines: 45
+- id: u2
+  summary: Update PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md IMP-18 row at L69 — flip status `pending`→`documented` + append gap-doc reference. Same-file atomic Edit.
+  files: [docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md]
+  tests: []
+  estimate_lines: 2
+
+Per-unit rationale (≤3 lines/unit):
+u1 — A1 (Phase R' source): `src/renderer.py:169-207` `_preprocess_svg_data`; `SVG_BLOCKS={"venn-diagram","relationship"}` L175; call site `src/renderer.py:321` in `render_multi_page()`; `src/svg_calculator.py:15-156` five helpers (`calc_circle_positions` L15, `calc_item_radius` L47, `calc_orbit_radius` L59, `calc_outer_radius` L70, `prepare_venn_data` L77). A2 (Phase Z partial surface): 13 families + 2 frames under `templates/phase_z2/`; grep `<svg|viewBox`=0 across 15 partials; F12 `construction_goals_three_circle_intersection.html` implements three-circle via HTML/CSS `border-radius:50%` + `::before` outer ring, not SVG. A3 (dependency): IMP-18 dormant until IMP-04 registers SVG-bearing partial. A4 (guardrail): `renderer.py` + `svg_calculator.py` remain unmodified (read-only references only).
+u2 — 2-line edit on L69: status `pending`→`documented` + append `docs/architecture/IMP-18-SVG-GAP-REPORT.md` reference. No anchor SHA bump (doc-only, no runtime source change). Same-file atomic Edit per workflow_atomicity_rules.
+
+Out-of-scope:
+- No edits to `src/renderer.py` / `src/svg_calculator.py` (Phase R' read-only guardrail).
+- No `<svg>` / `viewBox` insertion into the 15 Phase Z partials.
+- No port of `_preprocess_svg_data` into `phase_z2_renderer` (deferred until IMP-04 activation).
+- No new SVG-bearing frame_partial under `templates/phase_z2/` (owned by IMP-04).
+- No Phase Z 8-preset layout vocabulary / zone topology / fit-classifier change.
+- No F12 `construction_goals_three_circle_intersection` HTML→SVG migration.
+- `tests: []` per unit reflects doc-only deliverable (no runtime surface); pytest baseline captured in EVIDENCE.
+
+Rollback strategy:
+`git restore docs/architecture/IMP-18-SVG-GAP-REPORT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — no runtime surface, fixtures, or templates touched; single-command idempotent revert.
+
+=== EVIDENCE ===
+Commands run:
+- `ls templates/phase_z2/families/*.html | wc -l` → 13
+- `ls templates/phase_z2/frames/*.html | wc -l` → 2
+- Grep `<svg|viewBox` over `templates/phase_z2/` (ripgrep) → 0 matches across 15 partials
+- `test -f docs/architecture/IMP-18-SVG-GAP-REPORT.md` → NOT_EXISTS (u1 target clean)
+- Read `src/renderer.py:165-207` → `_preprocess_svg_data` def L169, `SVG_BLOCKS={"venn-diagram","relationship"}` L175, `prepare_venn_data` call L191, fallback warning L204, return L207
+- Read `src/renderer.py:315-329` (Round #1 baseline) → `_preprocess_svg_data(block_type, block_data)` call at L321 in `render_multi_page()`
+- Read `src/svg_calculator.py:1-156` (Round #1 baseline) → five helpers L15/L47/L59/L70/L77 with viewbox fields through L154
+- `rg -n "IMP-18" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → IMP-18 row at L69, status `pending`, soft link IMP-04 (matches scope-lock)
+- `git diff --stat HEAD -- src docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md templates/phase_z2` → empty (no in-flight drift on target surface)
+- `pytest -q tests` → 258 passed (Codex #1 Round #1 baseline; will re-run pre-Stage-3 verify)
+
+Files checked:
+- D:\ad-hoc\kei\design_agent\src\renderer.py (L165-329)
+- D:\ad-hoc\kei\design_agent\src\svg_calculator.py (L1-156)
+- D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md (L69)
+- D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html (13 files)
+- D:\ad-hoc\kei\design_agent\templates\phase_z2\frames\*.html (2 files)
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_simulation-plan_codex_r1.md

@@ -0,0 +1,56 @@
+[Codex #1] Stage 2 simulation-plan review (Round #1, post-rewind)
+
+Verdict: YES. The compact Stage 2 plan covers every issue-body axis and satisfies the post-rewind compact-plan contract.
+
+Plan completeness:
+- Axis 1, Phase R reference: covered by u1 via `src/renderer.py:169-207`, `src/renderer.py:321`, and `src/svg_calculator.py:15-156`.
+- Axis 2, Phase Z SVG precompute area: covered by u1 via 13 family partials + 2 frame partials under `templates/phase_z2/`, with zero `<svg>` / `viewBox` matches.
+- Axis 3, IMP-04 dependency: covered by u1 activation-gate language and u2 backlog cross-reference update.
+- Axis 4, Phase R guardrail: covered by doc-only scope and explicit no-edit scope for `src/renderer.py` / `src/svg_calculator.py`.
+
+Implementation unit check:
+- `=== IMPLEMENTATION_UNITS ===` exists.
+- u1: 1 file, 45 estimated lines, within <=3 files and <=50 lines.
+- u2: 1 file, 2 estimated lines, within <=3 files and <=50 lines.
+- Oversized unit: NO.
+
+Compact contract check:
+- The complete Claude draft is 4,668 chars, under the 5,000-char cap.
+- Body structure is limited to the implementation-units block, brief per-unit rationale, out-of-scope notes, rollback strategy, evidence, and final consensus marker.
+- No code snippets or inline code blocks are present.
+- The previous blocking `Axis coverage check` long-analysis section is absent.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-ChildItem -LiteralPath 'templates\\phase_z2\\families' -Filter '*.html' | Measure-Object` -> 13
+- `Get-ChildItem -LiteralPath 'templates\\phase_z2\\frames' -Filter '*.html' | Measure-Object` -> 2
+- `rg -n "<svg|viewBox" templates\\phase_z2` -> 0 matches, exit 1
+- Read `src/renderer.py:165-224` -> `_preprocess_svg_data` at L169, `SVG_BLOCKS` at L175, `prepare_venn_data` call at L191, fallback warning at L204, return at L207
+- Read `src/renderer.py:315-329` -> `_preprocess_svg_data(block_type, block_data)` call at L321 in `render_multi_page()` flow
+- Read `src/svg_calculator.py:1-154` -> helpers at L15, L47, L59, L70, L77 and returned viewbox fields through L154
+- `rg -n "IMP-18|SVG|frame_partials" docs\\architecture\\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` -> IMP-18 row at L69, status `pending`, soft link IMP-04
+- `git diff --stat HEAD -- src docs\\architecture\\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md templates\\phase_z2` -> empty
+- `Test-Path docs\\architecture\\IMP-18-SVG-GAP-REPORT.md` -> False
+- `(Get-Content -Raw .orchestrator\\drafts\\18_stage_simulation-plan_claude_r1.md).Length` -> 4668
+- `pytest -q tests` -> 258 passed in 40.80s
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\renderer.py`
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\frames\*.html`
+- `D:\ad-hoc\kei\design_agent\.orchestrator\drafts\18_stage_simulation-plan_claude_r1.md`
+
+Test results:
+- `pytest -q tests`: 258 passed in 40.80s.
+
+Verified facts:
+- Phase Z active partial inventory remains 13 family files and 2 frame files.
+- Existing Phase Z partials contain zero `<svg>` / `viewBox` matches.
+- IMP-18 backlog row is at L69 and currently `pending` before planned implementation.
+- Proposed gap-report path is absent before implementation, so u1 is a new-doc unit.
+- Scoped diff for source/docs/Phase Z templates is empty before implementation.
+- Phase R files remain read-only references in the plan.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_simulation-plan_codex_r2.md

@@ -0,0 +1,51 @@
+[Codex #2] Stage 2 simulation-plan review (Round #2, post-rewind)
+
+Verdict: YES. The Round #2 compact plan covers every issue-body axis and fixes the previous Stage 2 contract failure.
+
+Plan completeness:
+- Axis 1, Phase R reference: covered by u1 through `src/renderer.py:169-207`, `src/renderer.py:321`, and `src/svg_calculator.py:15-154`.
+- Axis 2, Phase Z SVG precompute area: covered by u1 through 13 family partials + 2 frame partials under `templates/phase_z2/`, with zero `<svg>` / `viewBox` matches.
+- Axis 3, IMP-04 dependency: covered by u1 activation-gate language and u2 backlog cross-reference update.
+- Axis 4, Phase R guardrail: covered by doc-only scope and explicit no-edit scope for `src/renderer.py` / `src/svg_calculator.py`.
+
+Implementation unit check:
+- `=== IMPLEMENTATION_UNITS ===` exists.
+- u1: 1 file, 45 estimated lines, within <=3 files and <=50 lines.
+- u2: 1 file, 2 estimated lines, within <=3 files and <=50 lines.
+- Oversized unit: NO.
+
+Compact contract check:
+- The full Claude Round #2 draft is 4,156 chars, under the 5,000-char cap.
+- The previous blocking long-analysis section is absent; no `Axis coverage`, `Plan completeness`, `Implementation unit check`, or `Compact contract` sections are present in the plan draft.
+- No fenced code blocks are present.
+- The plan body is limited to implementation units, brief per-unit rationale, out-of-scope notes, rollback strategy, evidence, and final consensus marker.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-ChildItem -LiteralPath 'templates\\phase_z2\\families' -Filter '*.html' | Measure-Object | Select-Object -ExpandProperty Count` -> 13
+- `Get-ChildItem -LiteralPath 'templates\\phase_z2\\frames' -Filter '*.html' | Measure-Object | Select-Object -ExpandProperty Count` -> 2
+- `rg -n "<svg|viewBox" templates\\phase_z2` -> 0 matches, exit 1
+- `Test-Path -LiteralPath 'docs\\architecture\\IMP-18-SVG-GAP-REPORT.md'` -> False
+- Read `src/renderer.py:165-224` -> `_preprocess_svg_data` at L169, `SVG_BLOCKS` at L175, `prepare_venn_data` call at L191, fallback warning at L204, return at L207
+- Read `src/renderer.py:315-329` -> `_preprocess_svg_data(block_type, block_data)` call at L321 in `render_multi_page()` flow
+- Read `src/svg_calculator.py:1-160` -> helpers at L15, L47, L59, L70, L77 and returned viewbox fields through L154
+- `rg -n "IMP-18|SVG|frame_partials" docs\\architecture\\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` -> IMP-18 row at L69, status `pending`, soft link IMP-04
+- `git diff --stat HEAD -- src docs\\architecture\\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md templates\\phase_z2` -> empty
+- `(Get-Content -Raw -LiteralPath '.orchestrator\\drafts\\18_stage_simulation-plan_claude_r2.md').Length` -> 4156
+- `rg -n "^```|Axis coverage|Plan completeness|Implementation unit check|Compact contract" .orchestrator\\drafts\\18_stage_simulation-plan_claude_r2.md` -> only the rollback inline command line, no forbidden analysis sections or fenced code blocks
+- `rg -n "construction_goals_three_circle_intersection|border-radius|::before|<svg|viewBox" templates\\phase_z2\\families\\construction_goals_three_circle_intersection.html` -> `border-radius: 50%`, `::before`, frame_id 1171281189; no SVG hits
+- `pytest -q tests` -> 258 passed in 40.42s
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\.orchestrator\drafts\18_stage_simulation-plan_claude_r2.md`
+- `D:\ad-hoc\kei\design_agent\src\renderer.py`
+- `D:\ad-hoc\kei\design_agent\src\svg_calculator.py`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\*.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\frames\*.html`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\families\construction_goals_three_circle_intersection.html`
+
+Test results:
+- `pytest -q tests` -> 258 passed in 40.42s
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_test-verify_claude_r1.md

@@ -0,0 +1,79 @@
+[Claude #3] Stage 4 test-verify Round #1 — IMP-18 SVG gap report doc-only carve-out
+
+## Verdict
+
+**PASS** — Stage 3 working-tree state (u1 + u2) matches Stage 2 plan; Phase R' read-only guardrail intact; Phase Z partial SVG inventory unchanged; doc-only scope → no runtime test surface.
+
+## Scope of this verification
+
+scope = Stage 3 working-tree diff vs. Stage 2 implementation_units; Phase R' guardrail; Phase Z partial inventory; gap-report axis coverage + line-ref accuracy; hardcoding check; anchor SHA bump applicability.
+NOT scope = full repo pytest, full repo grep, Phase R' source modification, IMP-04 readiness, IMP-18 re-activation.
+
+## Per-axis verification
+
+**Axis A1 — u1 deliverable (`docs/architecture/IMP-18-SVG-GAP-REPORT.md`, 64 lines, new file)**
+- A1 Phase R' source pattern citations cross-verified against working tree:
+  - `src/renderer.py:169` → `def _preprocess_svg_data(block_type, block_data)` — confirmed.
+  - `src/renderer.py:175` → `SVG_BLOCKS = {"venn-diagram", "relationship"}` — confirmed.
+  - `src/renderer.py:321` → `block_data = _preprocess_svg_data(block_type, block_data)` in `render_multi_page()` — confirmed.
+  - `src/svg_calculator.py:15` `calc_circle_positions` — confirmed.
+  - `src/svg_calculator.py:47` `calc_item_radius` — confirmed.
+  - `src/svg_calculator.py:59` `calc_orbit_radius` — confirmed.
+  - `src/svg_calculator.py:70` `calc_outer_radius` — confirmed.
+  - `src/svg_calculator.py:77` `prepare_venn_data` — confirmed.
+- A2 Phase Z partial inventory: filesystem 13 families + 2 frames = 15 partials, grep `<svg|viewBox` = 0 matches across all 15. (git ls-files shows 11 tracked; 4 untracked partials confirmed via filesystem glob — Stage 2 Get-ChildItem evidence used filesystem inventory, gap-report claim aligned with filesystem reality.)
+- A2 F12 evidence: `construction_goals_three_circle_intersection.html` HTML/CSS (`border-radius:50%` + `linear-gradient` + `::before`) not SVG — gap-report claim consistent with file content.
+- A3 IMP-04 dormant dependency stated explicitly with re-activation triggers (SVG-bearing partial under `templates/phase_z2/` + slot contract consumption).
+- A4 Phase R' guardrail enumerated for all 4 surfaces (`renderer.py`, `svg_calculator.py`, 15 phase_z2 partials, F12 migration scope exclusion).
+- Re-activation checklist appended (4 items) — forward-port guardrail consistent with A4.
+- File size: 64 lines vs. Stage 2 estimate 45 — variance due to 4-axis structure with line-level refs + re-activation checklist; within doc-only scope per Stage 2 exit `wc -l` evidence.
+
+**Axis A2 — u2 deliverable (`docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, L69, 1 insertion + 1 deletion)**
+- `git diff --stat` → `1 file changed, 1 insertion(+), 1 deletion(-)` — atomic same-file Edit confirmed.
+- `git diff` → only L69 IMP-18 row touched; status column `pending` → `documented`; description column appended `— [gap report](IMP-18-SVG-GAP-REPORT.md)`. No other rows altered.
+- workflow_atomicity_rules: same-file single-hunk edit — compliant.
+
+**Axis A3 — Phase R' read-only guardrail**
+- `git diff --stat src/renderer.py src/svg_calculator.py templates/phase_z2/` → empty output (no diff) — guardrail intact.
+- `git status --short src/renderer.py src/svg_calculator.py` → empty — no working-tree modifications.
+- Phase R' `_preprocess_svg_data` body, `SVG_BLOCKS` set, call site, and 5 svg_calculator helpers untouched.
+
+**Axis A4 — Hardcoding / anchor-sync / no-speculative checks**
+- Gap report cites file:line references, not hand-copied code snippets — no hardcoding.
+- No SVG coordinates introduced into Phase Z templates.
+- No anchor SHA bump applied; doc-only carve-out leaves no runtime source touched — appropriate per anchor_sync_rules ("anchor SHA tracks runtime contract changes, not doc-only").
+- No `_preprocess_svg_data` port into `phase_z2_renderer` (correctly deferred to IMP-04 activation).
+
+**Axis A5 — Test plan / regression**
+- Stage 2 plan: `tests: []` for both units (doc-only deliverable).
+- Manual verification per Stage 2 acceptance: `rg "<svg|viewBox" templates/phase_z2` = 0 (confirmed); `Test-Path docs/architecture/IMP-18-SVG-GAP-REPORT.md` = True (confirmed; 64 lines).
+- No runtime surface modified → pytest baseline parity is the expected post-condition; full pytest not run per "COMMIT SCOPE only" rule (working tree contains unrelated pre-existing modifications; running full pytest would mix scopes).
+
+## Out-of-scope confirmations
+
+- No port of `_preprocess_svg_data` into `phase_z2_renderer`.
+- No new SVG-bearing frame_partial under `templates/phase_z2/`.
+- No F12 HTML/CSS → SVG migration.
+- No Phase Z 8-preset layout vocabulary / zone topology / fit-classifier change.
+- No edits to `src/renderer.py` / `src/svg_calculator.py`.
+
+## Unresolved questions
+
+- Commit grouping: u1 (new doc) + u2 (1-line row flip) as single doc-only commit per "1 commit = 1 decision unit where decision unit = IMP-18 documented carve-out" — confirmed appropriate per workflow_atomicity_rules; final call deferred to Stage 5.
+- IMP-04 re-activation trigger surface — deferred to IMP-04 scope.
+
+=== EVIDENCE ===
+- `git status --short` (target paths) → ` M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` + `?? docs/architecture/IMP-18-SVG-GAP-REPORT.md`.
+- `git diff --stat docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → `1 file changed, 1 insertion(+), 1 deletion(-)`.
+- `git diff docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → L69 row: status `pending`→`documented`, appended `— [gap report](IMP-18-SVG-GAP-REPORT.md)`; no other rows altered.
+- `git diff --stat src/renderer.py src/svg_calculator.py templates/phase_z2/` → empty (guardrail intact).
+- `Glob templates/phase_z2/families/*.html` → 13 files (filesystem).
+- `Glob templates/phase_z2/frames/*.html` → 2 files (filesystem).
+- `Grep "<svg|viewBox" templates/phase_z2` → No files found (0 matches across 15 partials).
+- `Grep "SVG_BLOCKS|_preprocess_svg_data" src/renderer.py -n` → L169 def, L175 set definition, L177 guard, L321 call site — matches gap-report citations.
+- `Read src/renderer.py:165-214` → `_preprocess_svg_data` body confirms `SVG_BLOCKS = {"venn-diagram", "relationship"}` (L175), `prepare_venn_data` import (L189), exception fallback warning (L204).
+- `Read src/svg_calculator.py:1-160` → 5 helpers at L15/L47/L59/L70/L77 — matches gap-report citations; file ends at L156.
+- `Read docs/architecture/IMP-18-SVG-GAP-REPORT.md` → 64 lines, 4 axes (A1 source, A2 inventory, A3 dependency, A4 guardrail) + re-activation checklist.
+- HEAD = `e10ec36 feat(IMP-17): AI repair fallback infra carve-out — design-only boundary + 3-cond AND gate` (Stage 5 commit pending).
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/18_stage_test-verify_codex_r1.md

@@ -0,0 +1,73 @@
+[Codex #5] Stage 4 test-verify Round #1 - IMP-18 SVG gap report doc-only carve-out
+
+## Verdict
+
+PASS - Stage 3 working-tree state matches the Stage 2 plan for u1 + u2. The changed surface is doc-only, Phase R' read-only guardrails are intact, and the Phase Z partial SVG inventory remains unchanged.
+
+## Scope
+
+Verified scope = Stage 3 working-tree diff, u1/u2 deliverables, Phase R' guardrail, Phase Z SVG inventory, line-reference accuracy, backlog status update, and doc-only test applicability.
+
+Not verified = full pytest suite, IMP-04 readiness, future SVG partial onboarding, or any runtime renderer behavior.
+
+## Findings
+
+- u1 is present: `docs/architecture/IMP-18-SVG-GAP-REPORT.md` exists and has 64 lines.
+- u1 covers the required four axes: Phase R' source pattern, Phase Z 15-partial SVG absence, IMP-04 activation gate, and Phase R' read-only guardrail.
+- u2 is atomic: `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` has one hunk with one insertion and one deletion.
+- Backlog line 69 now marks IMP-18 as `documented` and appends `[gap report](IMP-18-SVG-GAP-REPORT.md)`.
+- `src/renderer.py`, `src/svg_calculator.py`, and `templates/phase_z2/` have no tracked diff for this issue.
+- Phase Z filesystem inventory is 13 family partials + 2 frame partials = 15 total partials.
+- `rg -n "<svg|viewBox" templates\phase_z2` returned no matches, exit code 1, which is expected for this no-SVG inventory check.
+- `construction_goals_three_circle_intersection.html` uses HTML/CSS geometry (`border-radius: 50%`, `linear-gradient`, and `::before`) and no SVG/viewBox match.
+
+## Test result
+
+No pytest run was required for this Stage 4 verification because the Stage 2 implementation units declare `tests: []` and the committed scope is documentation only. Verification used commit-scope diff and filesystem checks instead of full-suite runtime tests.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `git diff --stat -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `Get-Content -LiteralPath 'docs\architecture\IMP-18-SVG-GAP-REPORT.md'`
+- `$i=0; Get-Content -LiteralPath 'src\renderer.py' | ForEach-Object { ... lines 165-207 ... }`
+- `$i=0; Get-Content -LiteralPath 'src\renderer.py' | ForEach-Object { ... lines 316-323 ... }`
+- `$i=0; Get-Content -LiteralPath 'src\svg_calculator.py' | ForEach-Object { ... lines 15,47,59,70,77 ... }`
+- `(Get-ChildItem -LiteralPath 'templates\phase_z2\families' -Filter '*.html' | Measure-Object).Count; (Get-ChildItem -LiteralPath 'templates\phase_z2\frames' -Filter '*.html' | Measure-Object).Count`
+- `rg -n "<svg|viewBox" templates\phase_z2`
+- `git diff -- src/renderer.py src/svg_calculator.py templates/phase_z2`
+- `Select-String -LiteralPath 'templates\phase_z2\families\construction_goals_three_circle_intersection.html' -Pattern 'border-radius|linear-gradient|::before|<svg|viewBox' -CaseSensitive`
+- `@(Get-Content -LiteralPath 'docs\architecture\IMP-18-SVG-GAP-REPORT.md').Count`
+- `Test-Path -LiteralPath 'docs\architecture\IMP-18-SVG-GAP-REPORT.md'`
+- `Select-String -LiteralPath 'docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md' -Pattern 'IMP-18'`
+- `git diff --name-only -- docs/architecture/IMP-18-SVG-GAP-REPORT.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md src/renderer.py src/svg_calculator.py templates/phase_z2`
+- `git log --oneline -- docs/architecture/IMP-18-SVG-GAP-REPORT.md`
+- `git ls-files 'templates/phase_z2/families/*.html' 'templates/phase_z2/frames/*.html'`
+
+Files checked:
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `src/renderer.py`
+- `src/svg_calculator.py`
+- `templates/phase_z2/families/construction_goals_three_circle_intersection.html`
+- `templates/phase_z2/families/*.html`
+- `templates/phase_z2/frames/*.html`
+
+Test results:
+- Not run. Stage 2 planned `tests: []`; Stage 4 was verified with doc-only commit-scope checks.
+
+Verified facts:
+- `src/renderer.py:169` defines `_preprocess_svg_data`.
+- `src/renderer.py:175` defines `SVG_BLOCKS = {"venn-diagram", "relationship"}`.
+- `src/renderer.py:321` calls `_preprocess_svg_data` inside `render_multi_page()`.
+- `src/svg_calculator.py` contains the referenced helpers at lines 15, 47, 59, 70, and 77.
+- `docs/architecture/IMP-18-SVG-GAP-REPORT.md` exists and has 64 lines.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` line 69 is the IMP-18 row and now has status `documented`.
+- Backlog diff is exactly 1 insertion and 1 deletion.
+- Phase Z partial inventory is 13 families + 2 frames.
+- No `<svg>` or `viewBox` matches exist under `templates/phase_z2`.
+- No tracked diff exists in `src/renderer.py`, `src/svg_calculator.py`, or `templates/phase_z2`.
+- `git log --oneline -- docs/architecture/IMP-18-SVG-GAP-REPORT.md` is empty because the file is new and not yet committed.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_code-edit_claude_r1.md

@@ -0,0 +1,47 @@
+Executing unit: u1
+
+[Claude #6] Stage 3 code-edit Round #1 (post-rewind retry) — unit u1 corrected.
+
+**unit_executed**: u1 (post-rewind: A2 producer-trace factual fix per Codex #5 finding)
+
+**files_changed**:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` (still untracked; in-place correction of A1 L18 + A2 L40-44, no new files)
+
+**diff_summary**:
+- **Rewind driver**: Codex #5 (Stage 3 verifier) `FINAL_CONSENSUS: NO` / `rewind_target: stage_3_edit` — A2 falsely listed `font_size_px` and `padding_px` as Type B producer-emitted `block_constraints` keys; verified source shows they are renderer-side defaults at `src/renderer.py:262-263`, not producer output.
+- **A2 producer trace split** (was one bullet `src/space_allocator.py:541-588` claiming generic emission): replaced with three role-specific bullets reflecting verified source:
+  - `:541-556` top-role: `block_constraints = {"img_width_px": img_w, "img_height_px": top_h if img_w>0 else 0, "has_image": img_w>0}` — image-aware keys only.
+  - `:562-574` bottom-role: `block_constraints = {}` (empty; no producer keys).
+  - `:577-588` footer-role: `block_constraints = {}` (empty; `max_height_cost="low"` literal).
+- **A2 producer classification line (L44) corrected**: removed false `font_size_px` / `padding_px` from producer-output claim. New text explicitly states `block_constraints` contains *only* image-aware keys on top role and is *empty* on bottom/footer roles, and that `font_size_px` / `padding_px` are renderer-side defaults consumed via `.get(..., 15.2)` / `.get(..., 20)` at `src/renderer.py:262-263` — not producer output.
+- **A1 L18 clarified**: the inline-style injection bullet (`src/renderer.py:261-274`) now appends a note that `font_size` / `padding` are read at `:262-263` via `.get(..., default)` — renderer-side defaults, not producer-emitted (forward-pointer to A2). This makes the renderer-consumer vs producer-output distinction explicit per Codex #5 recommendation.
+- All references remain read-only file:line. No code body copying. No edits to `src/`.
+
+**guardrails honored (Stage 2 GR1-GR6)**:
+- GR1: No edits to `src/renderer.py`, `src/space_allocator.py`, `src/phase_z2_pipeline.py` — confirmed via `git status --short docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md src/` → only `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` is untracked, `src/` clean.
+- GR2: Docs-only, file:line references retained.
+- GR3: No role-based `["배경","본심"]` hardcoding injected — surfaced only as A1 reference text (unchanged).
+- GR4: A5 forward-pointer in header `Soft dependency` line preserved; A5 body still lands in u2.
+- GR5: IMP-09 boundary marker preserved in header; full A4 boundary still lands in u2.
+- GR6: `build_containers_type_b` call sites untouched.
+
+**verification (post-correction)**:
+- Re-read `src/space_allocator.py:445-594` — confirmed:
+  - top-role emission L541-556: `block_constraints` keys = {`img_width_px`, `img_height_px`, `has_image`}.
+  - bottom-role emission L562-574: `block_constraints = {}` at L573.
+  - footer-role emission L577-588: `block_constraints = {}` at L587, `max_height_cost="low"` literal at L586.
+- Re-read `src/renderer.py:255-279` — confirmed:
+  - L262: `font_size = spec.block_constraints.get("font_size_px", 15.2)` (renderer-side default `15.2`).
+  - L263: `padding = spec.block_constraints.get("padding_px", 20)` (renderer-side default `20`).
+- Re-read final doc `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` (45 lines) — A1 has 6 bullets + 2 call-site bullets + classification line; A2 has 9 bullets (was 8: one emission bullet split into three role-specific bullets) + classification line; header + Status/Scope/Soft-dependency/Source-axis unchanged.
+- `git status --short` confirms only `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` untracked; `src/` clean.
+- No pytest run — docs-only, runtime surface zero (Stage 2 test_plan unchanged).
+
+**remaining_units**:
+- u2: Append A3 (Phase Z solver delta — `src/phase_z2_pipeline.py:794-853` row + `:924-972` col + `:1125-1452` dispatch, IMP-09 owned) + A4 (IMP-09 boundary statement) + A5 (re-activation gate + GR1-GR4) to `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`.
+- u3: Update `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — IMP-19 row append doc link + status pending → documented; IMP-09 row append back-reference.
+- u4: Update `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 I4 row (line 125) — prepend IMP-19 ID anchor + reference doc link.
+
+**follow_up_issue_candidates**: none. The producer-trace correction stays within u1 scope; no new axis observed. Stage 2 plan remains valid; rewind consumed only one round.
+
+STOP — no commit, no push, no anticipation of u2. Per RULE PZ-2 (1 turn = 1 step) and Stage 3 contract (one unit per turn). Awaiting Codex verifier review.

.orchestrator/drafts/19_stage_code-edit_claude_r2.md

@@ -0,0 +1,40 @@
+Executing unit: u2
+
+[Claude #7] Stage 3 code-edit Round #2 — unit u2 executed.
+
+**unit_executed**: u2 (append A3 + A4 + A5 to `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`)
+
+**files_changed**:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` (still untracked; appended A3 + A4 + A5 sections after A2)
+
+**diff_summary**:
+- Appended **A3 — Phase Z Step 8 solver delta (IMP-09 owned)** with verified file:line refs:
+  - `src/phase_z2_pipeline.py:794-853` `compute_zone_layout` row solver — algorithm = `min_height_first + content_weight_distribution`; Step 1 reserves per-zone `min_height_px` from frame_contract `visual_hints` (with proportional scale-down on overflow), Step 2 distributes remaining vertical budget by `content_weight.score`, Step 3 absorbs rounding residual into last zone.
+  - `src/phase_z2_pipeline.py:924-972` `compute_zone_layout_cols` col solver — algorithm = `content_weight_distribution_cols` (weight-only; no `min_width_px` contract per IMP-09 verification); zero-weight guard splits evenly.
+  - `src/phase_z2_pipeline.py:1125-1452` topology dispatch surface — `_build_rows_dynamic` (rows / horizontal-2), `_build_grid_dynamic_2d` (T / inverted-T / side-T-left / side-T-right / 2x2), dynamic-branch dispatcher at `:1444-1452`, user-override branch at `:1380-1434` (`computation == "user_override_geometry"`).
+  - Delta-vs-Phase-O table on 7 axes (geometry level / width source / height source / role axis / min-height source / topology dispatch / inline-style injection) to surface the abstraction-layer separation that justifies "no direct integration".
+- Appended **A4 — IMP-09 boundary statement (soft-link)**: explicit ownership split — IMP-09 owns solver algorithm + dispatch surface + `visual_hints.min_height_px` contract; IMP-19 owns reference-only documentation. No bidirectional code flow. Reference direction is one-way (docs → src/, never reverse). If IMP-09 alters solver signature, A3 file:line refs must be re-verified; boundary itself does not move.
+- Appended **A5 — Re-activation gate + guardrails**: 4-condition gate (trigger = Phase Z solver regression mapping to block-level role abstraction; evidence = failing MDX + frame_contract trace + geometry; IMP-09 sign-off; scope re-lock under fresh issue). Guardrails GR1-GR4 inline-stated as durable doc-side contracts (GR1 no runtime integration, GR2 Phase O no-regression, GR3 reference extract stays in `docs/architecture/`, GR4 soft-link integrity).
+- File grew from 45 → 97 lines (delta = 52 lines; Stage 2 estimate was 36; overage came from the 7-row delta-comparison table in A3 + GR1-GR4 inline restatement, both content-additive — no scope creep).
+- All references remain read-only `file:line` only. No code body copying. No edits to `src/`.
+
+**guardrails honored (Stage 2 GR1-GR6)**:
+- GR1: No edits to `src/renderer.py`, `src/space_allocator.py`, `src/phase_z2_pipeline.py` — confirmed via `git status --short -- src/ docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` → only `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` is untracked; `src/` clean.
+- GR2: Docs-only, file:line references only; A3 delta table is structural narrative, not code copy.
+- GR3: No role-based `["배경","본심"]` hardcoding injected into Phase Z; A3 mentions it only as a Phase O reference axis to highlight the absence of role concept in Phase Z.
+- GR4: A5 gate fully expressed; IMP-19 stays `documented` (dormant) until gate fires.
+- GR5: IMP-09 boundary statement (A4) explicit; IMP-19 does **not** absorb solver ownership.
+- GR6: `build_containers_type_b` call sites untouched.
+
+**verification (A3 file:line refs)**:
+- Re-read `src/phase_z2_pipeline.py:790-853` — confirmed `compute_zone_layout` signature + 3-step algorithm + return keys (`heights_px`, `ratios`, `computation == "min_height_first + content_weight_distribution"`).
+- Re-read `src/phase_z2_pipeline.py:920-972` — confirmed `compute_zone_layout_cols` signature + weight-only algorithm + zero-weight guard + return keys (`widths_px`, `width_ratios`, `computation == "content_weight_distribution_cols"`).
+- Re-read `src/phase_z2_pipeline.py:1120-1180` — confirmed `_build_rows_dynamic` (topology=="rows") + `_build_grid_dynamic_2d` (T / inverted-T / side-T-left / side-T-right / 2x2) entry points.
+- Re-read `src/phase_z2_pipeline.py:1380-1452` — confirmed user-override geometry branch (`computation == "user_override_geometry"`) + dynamic-branch dispatcher at `:1444-1452`.
+- No factual claim made about producer-side `font_size_px` / `padding_px` (post-rewind correctness preserved from u1).
+
+**remaining_units**: [u3, u4]
+- u3 = `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — IMP-19 row append doc link + status `pending` → `documented`; IMP-09 row append back-reference to IMP-19 doc (~4 lines).
+- u4 = `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 I4 row (line 125) — prepend IMP-19 ID anchor + reference doc link (~2 lines).
+
+**follow_up_issue_candidates**: none — A3 surface kept strictly within IMP-09 reference scope; A4 + A5 do not introduce new axes. The A5 gate is a documentation contract, not a deferred work item.

.orchestrator/drafts/19_stage_code-edit_claude_r3.md

@@ -0,0 +1,36 @@
+Executing unit: u3
+
+[Claude #8] Stage 3 code-edit Round #3 — unit u3 executed.
+
+**unit_executed**: u3 (PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md — IMP-19 row append doc link + status pending → documented; IMP-09 row append back-reference to IMP-19 doc)
+
+**files_changed**:
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (modified — 2 row updates: IMP-09 dependency column + IMP-19 scope column & status column)
+
+**diff_summary**:
+- **IMP-19 row (line 70)** — `pending` → `documented`; scope column appended ` — [reference doc](IMP-19-ZONE-RATIO-REFERENCE.md)` (mirrors IMP-18 precedent line 69 `... — [gap report](IMP-18-SVG-GAP-REPORT.md) ... | documented`).
+- **IMP-09 row (line 53)** — dependency column `none` → `soft back-link: IMP-19 ([reference doc](IMP-19-ZONE-RATIO-REFERENCE.md) — Phase O block-level pattern reference, no runtime integration)`. The `soft back-link` term distinguishes the reverse pointer from forward `soft link` usage in the same column (IMP-19 row still keeps forward `soft link: IMP-09`). The parenthetical explicitly restates the "no runtime integration" boundary from A4 of the reference doc to prevent future readers from mis-reading the back-link as a runtime dependency.
+- IMP-09 `status = implemented` preserved (back-link does not alter implementation state; IMP-09 solver is shipped per IMP-09 Stage 6).
+- No edits to any other row, no edits to header / footnote / §3 / §4 sections of the backlog. Diff confirmed via `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` = exactly 2 line replacements (one per row), zero adjacent-row drift.
+- Stage 2 estimate for u3 was 4 lines; actual = 2 changed lines (each row is a single-line markdown table row, so the 4-line estimate counted the two old + two new states). No overage, no scope creep.
+
+**guardrails honored (Stage 2 GR1-GR6)**:
+- GR1: No edits to `src/renderer.py`, `src/space_allocator.py`, `src/phase_z2_pipeline.py` — confirmed via `git status --short -- docs/architecture/ src/` → only `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` modified, `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` untracked (from u1+u2), `src/` clean.
+- GR2: Docs-only, file:line references retained (no code body copying).
+- GR3: No role-based `["배경","본심"]` hardcoding injected — IMP-19 scope text unchanged on this axis (only `— [reference doc](...)` link appended).
+- GR4: Soft-link integrity preserved — IMP-19 → IMP-09 forward `soft link` unchanged, IMP-09 → IMP-19 reverse `soft back-link` added with explicit "no runtime integration" disambiguator.
+- GR5: IMP-09 boundary preserved — `status = implemented` and scope column ("`build_layout_css` 분기 확장…") untouched; solver ownership not transferred to IMP-19.
+- GR6: `build_containers_type_b` call sites untouched (the term appears only in the IMP-09 guardrail column text, which is unchanged).
+
+**verification**:
+- `git status --short -- docs/architecture/ src/` → ` M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` + `?? docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`. No `src/` files modified.
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` → 2 row-replacement hunks (line 53 IMP-09 dependency; line 70 IMP-19 scope+status). Adjacent rows (IMP-06/07/08, IMP-10/11, IMP-16/17/18, IMP-20) all present unchanged in diff context, confirming no drift.
+- IMP-19 row now matches IMP-18 documentation pattern (`§3 Reference Only` + reference doc link in scope + `documented` status).
+- Reference doc link `IMP-19-ZONE-RATIO-REFERENCE.md` resolves to the file created in u1+u2 (still untracked, same directory as backlog — relative link valid).
+- `pytest` not run — u3 is docs-only, runtime surface zero (Stage 2 test_plan: "docs-only, runtime surface = zero. No new tests required").
+
+**remaining_units**: [u4]
+
+**follow_up_issue_candidates**: none — u3 stayed strictly inside Stage 2 scope-lock. The `soft back-link` term is a doc-side annotation, not a new axis. If a future axis wants to formalize reverse-link semantics across all backlog rows, that would be a separate INTEGRATION-AUDIT-style issue, not a u3 follow-up.
+
+DO NOT commit or push per Stage 3 contract. STOP.

.orchestrator/drafts/19_stage_code-edit_claude_r4.md

@@ -0,0 +1,47 @@
+Executing unit: u4
+
+[Claude #9] Stage 3 code-edit Round #4 — unit u4 executed.
+
+**unit_executed**: u4 (PHASE-Q-INSIGHT-TO-22STEP-MAP.md §3 I4 row — prepend IMP-19 ID anchor + reference doc link)
+
+**files_changed**:
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (modified — exactly 1 line replacement at §3 axis table line 125)
+
+**diff_summary**:
+- **§3 axis table line 125 (I4 row)** — prepended `IMP-19 ` to the axis name and appended ` (reference — see [`IMP-19-ZONE-RATIO-REFERENCE.md`](IMP-19-ZONE-RATIO-REFERENCE.md))` to the axis column. Other 4 columns (step / source / status / normal-path) untouched per Stage 2 plan ("No other row content change").
+- **Before**: `| I4 zone 비중 분배 | Step 8 | §2.8 I4 (`renderer._group_blocks_by_area`) | pending | yes (deterministic) |`
+- **After**:  `| IMP-19 I4 zone 비중 분배 (reference — see [`IMP-19-ZONE-RATIO-REFERENCE.md`](IMP-19-ZONE-RATIO-REFERENCE.md)) | Step 8 | §2.8 I4 (`renderer._group_blocks_by_area`) | pending | yes (deterministic) |`
+- **Pattern mirrors IMP-17 precedent (line 123)** — `IMP-17 AI repair fallback infra (carve-out — see [`IMP-17-CARVE-OUT.md`](IMP-17-CARVE-OUT.md))` → same `IMP-XX <axis name> (<kind> — see [doc.md](doc.md))` form. The discriminator word for IMP-19 is `reference` (vs IMP-17 `carve-out`) because IMP-19 is reference-only / dormant per A5 gate (no runtime activation), while IMP-17 is an active carve-out for AI fallback infra. The discriminator is content-meaningful, not cosmetic.
+- **`status = pending` deliberately preserved** — this column in §3 tracks *axis runtime activation*, not *documentation status*. IMP-17 also remains `pending` here despite its carve-out doc existing; same semantic for IMP-19 (dormant per A5 re-activation gate). The backlog row (u3) carries the doc-tracking `documented` status; §3 here carries the runtime-activation `pending` status. Two columns in two files, two different axes — no sync mismatch.
+- Stage 2 estimate for u4 was 2 lines (old + new states counted); actual diff = exactly 1 line replacement (1 deletion + 1 insertion). No overage, no scope creep.
+- Diff confirmed via `git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` = exactly one `-`/`+` pair at L125, zero adjacent-row drift.
+- LF→CRLF git warning surfaced as expected on Windows checkout; no content impact.
+
+**guardrails honored (Stage 2 GR1-GR6)**:
+- GR1: No edits to `src/renderer.py`, `src/space_allocator.py`, `src/phase_z2_pipeline.py` — confirmed via `git status --short -- docs/architecture/ src/`:
+  - ` M docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (u4 this round)
+  - ` M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (u3 prior round, accepted)
+  - `?? docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` (u1+u2 prior rounds, accepted)
+  - `src/` clean.
+- GR2: Docs-only. The only file-level reference added is `renderer._group_blocks_by_area` (already present in original line) — no new code body copying.
+- GR3: No role-based `["배경","본심"]` hardcoding injected — the I4 row text refers to `renderer._group_blocks_by_area` symbol only, not its role-loop internals.
+- GR4: Soft-link integrity preserved — IMP-19 anchor now lands in 3 places (reference doc itself, backlog IMP-19 row + IMP-09 back-link row from u3, and §3 I4 row from u4). All three point to the same `IMP-19-ZONE-RATIO-REFERENCE.md` artifact. No bidirectional code flow introduced.
+- GR5: IMP-09 boundary preserved — §3 I4 axis remains tagged at Step 8 with `renderer._group_blocks_by_area` (Phase O block-level pattern). IMP-09 solver ownership (`phase_z2_pipeline.py:794-853 / :924-972 / :1125-1452`) untouched in this row.
+- GR6: `build_containers_type_b` call sites untouched (the symbol does not appear in this §3 row at all; it lives in A1+A2 of the reference doc only).
+
+**verification**:
+- `git status --short -- docs/architecture/ src/` (above) — scope confirmed.
+- `git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` — exactly 1 line replacement at L125. No header / §1 / §2 / §4 / §5 drift. The 3 surrounding rows (`IMP-17 AI repair fallback infra ...`, `I3 SVG 좌표 보강 ...`, `H2 frame contract validation ...`) are byte-identical in pre/post diff context.
+- Re-read of `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:115-127` confirmed the §3 axis table header (`axis | 관련 step | source | status | normal path 여부`) intact and I4 row now carries the IMP-19 anchor as the first table column.
+- Re-confirmed that `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` (u1+u2 artifact) lives at the path the new anchor points to — link target resolves.
+
+**unit_executed (id)**: u4
+
+**files_changed (list)**:
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+
+**remaining_units**: [] — all 4 Stage 2 implementation_units (u1, u2, u3, u4) are now executed. Codex r1/r2/r3 already verified u1+u2+u3. u4 awaits this round's verifier review.
+
+**follow_up_issue_candidates**: none. The Stage 1 / Stage 2 scope-lock for IMP-19 (docs-only, reference-only, soft-link to IMP-09) is fully satisfied by u1+u2+u3+u4. No scope-creep axes surfaced during execution. The A5 re-activation gate (in `IMP-19-ZONE-RATIO-REFERENCE.md`) defines the conditions under which a fresh issue would re-open this axis as a runtime concern; until those conditions trigger, IMP-19 stays `documented` (backlog) / `pending` (§3 axis runtime activation), as designed.
+
+STOP. Awaiting Codex r4 review.

.orchestrator/drafts/19_stage_code-edit_codex_r1.md

@@ -0,0 +1,37 @@
+[Codex #N]
+
+Stage 3 verifier review for the single executed unit `u1` (post-rewind retry): accepted.
+
+The correction fixes the prior factual producer-trace error. The new A2 text now distinguishes Type B producer output from renderer-side default consumption: `build_containers_type_b` emits image-aware `block_constraints` only for the top role, emits empty `block_constraints` for bottom/footer roles, and does not produce `font_size_px` or `padding_px`. The A1 consumer note correctly identifies those two keys as renderer `.get(..., default)` reads at `src/renderer.py:262-263`.
+
+Scope is correct for this unit: the only issue-related implementation file in the checked path set is the new docs artifact, and `src/renderer.py`, `src/space_allocator.py`, and `src/phase_z2_pipeline.py` were not modified. No full pytest was run because the current Stage 3 task is limited to verifying the single executed docs-only unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `git diff -- docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `Get-Content -LiteralPath docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- line-window read of `src/space_allocator.py:445-595`
+- line-window read of `src/renderer.py:210-296`
+- line-window read of `src/renderer.py:348-354` and `src/renderer.py:422-428`
+- `Select-String` on `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` for `font_size_px`, `padding_px`, `block_constraints`, and cited producer ranges
+- `git status --short -- src/renderer.py src/space_allocator.py src/phase_z2_pipeline.py docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+
+Files checked:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `src/space_allocator.py`
+- `src/renderer.py`
+- `src/phase_z2_pipeline.py` (status-only guardrail check)
+
+Test results: not run; Stage 3 requested single-unit verification only, and `u1` is docs-only.
+Remaining units: [u2, u3, u4]
+Verified facts:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` A1 cites `src/renderer.py:261-274` and explicitly says `font_size_px` / `padding_px` are renderer-side defaults read via `.get(..., 15.2)` / `.get(..., 20)` at `src/renderer.py:262-263`, not producer-emitted.
+- `src/space_allocator.py:541-556` top-role `ContainerSpec.block_constraints` contains `img_width_px`, `img_height_px`, and `has_image` only.
+- `src/space_allocator.py:562-574` bottom-role `ContainerSpec.block_constraints` is `{}`.
+- `src/space_allocator.py:577-588` footer-role `ContainerSpec.block_constraints` is `{}`.
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` A2 now matches those three producer cases and no longer claims Type B emits `font_size_px` or `padding_px`.
+- `src/renderer.py:352-353` passes `layout_concept.get("_container_specs")` into `_group_blocks_by_area`; `src/renderer.py:426` calls `_group_blocks_by_area(blocks_raw)` without `container_specs`.
+- Targeted status check shows only `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` under the issue path set; no checked `src/` file is modified.
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/19_stage_code-edit_codex_r2.md

@@ -0,0 +1,44 @@
+[Codex #N]
+
+Stage 3 verifier review for the single executed unit `u2`: accepted.
+
+The appended A3/A4/A5 material matches the Stage 2 plan and stays inside the docs-only boundary. A3 correctly describes the active Phase Z row solver as `min_height_first + content_weight_distribution`, the column solver as `content_weight_distribution_cols`, and the topology dispatch surface through rows, cols, 2-D presets, override geometry, and default fr fallback. A4 preserves the IMP-09 ownership boundary, and A5 documents the dormant re-activation gate plus GR1-GR4 without authorizing runtime integration.
+
+Scope is correct for this unit: the only issue-related implementation file in the checked path set is `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`. No checked `src/` file is modified. No full pytest was run because the current Stage 3 task is limited to verifying the single executed docs-only unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short -- src/renderer.py src/space_allocator.py src/phase_z2_pipeline.py docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `git diff -- docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `Get-Content -LiteralPath docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- line-window read of `src/phase_z2_pipeline.py:794-853`
+- line-window read of `src/phase_z2_pipeline.py:924-972`
+- line-window read of `src/phase_z2_pipeline.py:1125-1452`
+- `rg -n "min_width|min_width_px|visual_hints|min_height_px|frame_contracts" .`
+- `rg -n "def _build_fr_default|def _aggregate_zone_signals_per_track|_build_fr_default|_build_cols_dynamic" src/phase_z2_pipeline.py`
+- `rg -n "min_width_px|min_width:" templates/phase_z2/catalog/frame_contracts.yaml tests/phase_z2/test_compute_zone_layout_cols.py src/phase_z2_pipeline.py`
+- `Select-String` on `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` for A3/A4/A5, solver refs, IMP-09, and GR1-GR4
+
+Files checked:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `src/phase_z2_pipeline.py`
+- `templates/phase_z2/catalog/frame_contracts.yaml` (searched for `min_width_px` / `min_width`)
+- `tests/phase_z2/test_compute_zone_layout_cols.py` (searched for the no-`min_width_px` contract note)
+- `src/renderer.py` (status-only guardrail check)
+- `src/space_allocator.py` (status-only guardrail check)
+
+Test results: not run; Stage 3 requested single-unit verification only, and `u2` is docs-only.
+Remaining units: [u3, u4]
+Verified facts:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` now contains A3, A4, and A5 after the already accepted A1/A2 sections.
+- `src/phase_z2_pipeline.py:794-853` defines `compute_zone_layout`; it reserves `min_height_px`, scales minima down when total minimum height exceeds available height, distributes remaining height by `content_weight.score`, absorbs rounding residual into the last zone, and returns `heights_px`, `ratios`, and trace fields.
+- `src/phase_z2_pipeline.py:924-972` defines `compute_zone_layout_cols`; it distributes available width by `content_weight.score`, has a zero-weight even-split guard, and returns `widths_px` plus `width_ratios`.
+- The column no-`min_width_px` claim is supported by `src/phase_z2_pipeline.py:930-932`, `tests/phase_z2/test_compute_zone_layout_cols.py:4`, and no `min_width_px` / `min_width` hits in `templates/phase_z2/catalog/frame_contracts.yaml`.
+- `src/phase_z2_pipeline.py:1125-1452` includes `_build_rows_dynamic`, `_build_grid_dynamic_2d`, `_build_cols_dynamic`, override-geometry handling, dynamic dispatch for `rows`, `cols`, `T`, `inverted-T`, `side-T-left`, `side-T-right`, `2x2`, and `_build_fr_default` fallback.
+- A3's delta table correctly keeps Phase O/Q role-container behavior separate from Phase Z zone-level geometry.
+- A4 correctly keeps solver and topology ownership under IMP-09 while limiting IMP-19 to reference-only documentation.
+- A5 correctly leaves IMP-19 dormant and requires a future trigger, evidence, IMP-09 sign-off, and fresh scope lock before re-activation.
+- Targeted status check shows only `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` under the issue path set; no checked `src/` file is modified.
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/19_stage_code-edit_codex_r3.md

@@ -0,0 +1,33 @@
+[Codex #N]
+
+Stage 3 verifier review for the single executed unit `u3`: accepted.
+
+The backlog update matches the Stage 2 plan. The IMP-19 row now links to `IMP-19-ZONE-RATIO-REFERENCE.md` and moves from `pending` to `documented`. The IMP-09 row adds a soft back-link to the same reference doc while preserving `implemented` status and explicitly stating that the reference is a Phase O block-level pattern reference with no runtime integration.
+
+Scope is correct for this unit: the diff is limited to `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, with the previously accepted `IMP-19-ZONE-RATIO-REFERENCE.md` still untracked from u1/u2. No checked `src/` file is modified. No full pytest was run because the current Stage 3 task is limited to verifying the single executed docs-only unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short -- docs/architecture/ src/`
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `Select-String -LiteralPath docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md -Pattern 'IMP-09|IMP-19|IMP-18' -Context 0,0`
+- `Select-String -LiteralPath docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md -Pattern 'A4|IMP-09|no runtime integration|reference doc|A5|GR1|GR4' -Context 0,1`
+
+Files checked:
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `src/` status guardrail check
+
+Test results: not run; Stage 3 requested single-unit verification only, and `u3` is docs-only.
+Remaining units: [u4]
+Verified facts:
+- The IMP-09 row dependency column changed from `none` to `soft back-link: IMP-19 ([reference doc](IMP-19-ZONE-RATIO-REFERENCE.md) - Phase O block-level pattern reference, no runtime integration)`.
+- The IMP-09 row keeps `status = implemented`; its scope and guardrail columns are unchanged.
+- The IMP-19 row scope column now appends `[reference doc](IMP-19-ZONE-RATIO-REFERENCE.md)`.
+- The IMP-19 row status changed from `pending` to `documented`; its dependency remains `soft link: IMP-09`.
+- The IMP-18 row remains a matching documented-reference precedent and was not modified.
+- `IMP-19-ZONE-RATIO-REFERENCE.md` A4 states the IMP-09 ownership boundary and no bidirectional code flow; A5 keeps IMP-19 documented until the re-activation gate fires.
+- Targeted status check shows no checked `src/` file modified.
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/19_stage_code-edit_codex_r4.md

@@ -0,0 +1,34 @@
+[Codex #N]
+
+Stage 3 verifier review for the single executed unit `u4`: accepted.
+
+The insight-map update matches the Stage 2 plan. The §3 I4 row now carries the `IMP-19` anchor and links to `IMP-19-ZONE-RATIO-REFERENCE.md` in the same pattern as the nearby IMP-17 documented-axis precedent. The row still points at Step 8 and `§2.8 I4 (renderer._group_blocks_by_area)`, and it preserves `status = pending`, which is consistent with the reference doc's dormant A5 gate and the prior IMP-17 row semantics in this table.
+
+Scope is correct for this unit: the diff for `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` is exactly one markdown table row replacement. The previously accepted `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` modification and untracked `IMP-19-ZONE-RATIO-REFERENCE.md` artifact remain present from u1-u3. No checked `src/` file is modified. No full pytest was run because the current Stage 3 task is limited to verifying the single executed docs-only unit.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short -- docs/architecture/ src/`
+- `git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `Select-String -LiteralPath docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md -Pattern 'IMP-17|IMP-19|I4 zone|renderer._group_blocks_by_area' -Context 0,0`
+- `Select-String -LiteralPath docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md -Pattern 'IMP-09|IMP-19' -Context 0,0`
+- `Select-String -LiteralPath docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md -Pattern 'A5|re-activation|GR1|IMP-09|runtime integration' -Context 0,1`
+
+Files checked:
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `src/` status guardrail check
+
+Test results: not run; Stage 3 requested single-unit verification only, and `u4` is docs-only.
+Remaining units: none
+Verified facts:
+- The `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` diff is limited to one `-`/`+` pair at the I4 axis row.
+- The I4 row changed from `I4 zone 비중 분배` to `IMP-19 I4 zone 비중 분배 (reference — see [IMP-19-ZONE-RATIO-REFERENCE.md](IMP-19-ZONE-RATIO-REFERENCE.md))`.
+- The same row still has `Step 8`, `§2.8 I4 (renderer._group_blocks_by_area)`, `pending`, and `yes (deterministic)` in the remaining columns.
+- The nearby IMP-17 row already uses the same documented-axis pattern with an `IMP-17` anchor and a parenthesized doc link.
+- The IMP-19 backlog row remains `documented` with a reference-doc link, while the insight-map row remains `pending`; this is not a mismatch because the backlog tracks documentation state and the insight-map axis table tracks runtime activation.
+- `IMP-19-ZONE-RATIO-REFERENCE.md` A5 says IMP-19 is dormant/documented until the re-activation gate fires, and GR1 says the document does not authorize runtime integration.
+- Targeted status check shows no checked `src/` file modified.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_commit-push_claude_r1.md

@@ -0,0 +1,97 @@
+📌 **[오케스트레이터] Stage 5: 커밋 및 푸쉬 R1 — [Claude #5]**
+
+■ 핵심 결정 (Korean)
+- IMP-19 docs-only carve-out 4 units (u1+u2 신규 reference doc 97 lines + u3 backlog 2-row 갱신 + u4 INSIGHT-MAP §3 I4 row 1-hunk) 단일 커밋 `e60aacc` 으로 lock 완료.
+- 스테이징 = **정확히 3 파일만** (Stage 2 plan u1-u4 와 1:1 일치). 다른 모든 working-tree 변경 (orchestrator.py / chromedriver / matching/ 산출물 / samples / templates / scripts 등) 은 의도적으로 untouched 유지.
+- Push = `slide2` (Gitea `C.E.L_Slide_test2` — issue #19 host) + `origin` (GitHub `keimin86/design_agent`) 양쪽 모두 성공. 두 remote 의 `refs/heads/main` 모두 `e60aacc` 로 일치.
+- 런타임 행동 변경 = 0. `src/` delta = 0 (`renderer.py` / `space_allocator.py` / `phase_z2_pipeline.py` 모두 0-diff). 본 commit 은 순수 docs-only.
+- 본 commit 으로 IMP-19 가 backlog 에서 `documented` 상태로 영구 귀속됨 — IMP-09 owner 가 A5 gate 의 4 조건을 모두 충족하기 전까지 dormant reference axis.
+
+■ 범위 제외
+- `src/renderer.py` (Phase O `_group_blocks_by_area` 소스) / `src/space_allocator.py` (Phase O `build_containers_type_b` 소스) / `src/phase_z2_pipeline.py` (Phase Z Step 8 solver — IMP-09 owned) 수정 — Stage 1 GR1 read-only guardrail.
+- `role_order = ["배경", "본심"]` 패턴의 Phase Z 런타임 주입 — A5 gate + Stage 1 lock.
+- `compute_zone_layout` / `compute_zone_layout_cols` solver 알고리즘 교체 — IMP-09 표면.
+- `ContainerSpec.block_constraints` 의 Phase Z frame_contract 이관 — 별 axis.
+- `build_containers_type_b` call-site 변경 — Stage 2 GR6.
+- 신규 pytest 작성 / 기존 테스트 수정 — docs-only, runtime surface 0.
+- working tree 의 다른 변경 (orchestrator.py / matching/ 산출물 / chromedriver / samples / scripts / templates/catalog 등) — 별 axis 책임, 본 커밋에 비포함.
+
+■ 다음 단계
+- Stage 6 (이슈 종료 + final report). Codex 양측 검증 후 issue #19 close.
+
+=== EXIT REPORT (English, binding contract) ===
+
+**commit_sha**: `e60aacc` (full: `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`)
+
+**parent_sha**: `02e2ae0` (HEAD at Stage 1/2/3/4 entry — `docs(#54): F-4 legacy annotation + F-5 fixture convention -- AUDIT-01 housekeeping`)
+
+**files_committed** (3 files, +100 / −3):
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` — new, +97 lines (5 axes: A1 Phase O/Q consumer `_group_blocks_by_area` refs / A2 Phase O upstream producer `build_containers_type_b` refs / A3 Phase Z Step 8 solver delta + 7-row comparison table / A4 IMP-09 boundary statement / A5 re-activation gate + GR1-GR4).
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — +2 / −2 (row L53 IMP-09 dependency cell: `none` → `soft back-link: IMP-19 (...reference doc...)`; row L70 IMP-19: status `pending` → `documented` + `[reference doc](IMP-19-ZONE-RATIO-REFERENCE.md)` link).
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` — +1 / −1 (row L125 §3 I4 cell: `I4 zone 비중 분배` → `IMP-19 I4 zone 비중 분배 (reference — see [IMP-19-ZONE-RATIO-REFERENCE.md](...))`; step/classification preserved).
+
+**files_excluded** (intentionally NOT staged — out of IMP-19 scope, untouched in working tree):
+- `src/renderer.py`, `src/space_allocator.py`, `src/phase_z2_pipeline.py` — Phase O consumer / Phase O producer / Phase Z Step 8 solver references (Stage 1 GR1 + Stage 2 plan).
+- `orchestrator(원본).py`, `tests/PIPELINE.md`, `tests/PLAN.md`, `tests/PROGRESS.md`, `tests/README.md`, `tests/matching/*` — separate axes / pre-existing untracked artifacts.
+- `chromedriver/win64/147.0.7727.117/`, `samples/uploads/`, `samples/mdx/05*.mdx`, `samples/mdx_batch/05.mdx`, `scripts/post_gitea_comment.py`, `scripts/sync_phase_z_run_to_frontend.py` — separate axes.
+- `templates/catalog/`, `templates/phase_z2/families/app_sw_package_vs_solution.html`, `templates/phase_z2/families/pre_construction_model_info_stacked.html`, `templates/phase_z2/frames/` — separate axes.
+- `.orchestrator/`, `.claude/settings.local.json` — local-only.
+
+**push_targets**:
+- `slide2` (Gitea `Kyeongmin/C.E.L_Slide_test2` — issue #19 host) → `refs/heads/main` = `e60aacc` ✅
+- `origin` (GitHub `keimin86/design_agent`) → `refs/heads/main` = `e60aacc` ✅
+- `gitea` (Gitea `Kyeongmin/_Geulbeot`) — NOT pushed (separate project per memory `reference_design_agent_remotes`).
+
+**verification**:
+- `git rev-parse HEAD` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8` ✓
+- `git ls-remote slide2 main` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8 refs/heads/main` ✓
+- `git ls-remote origin main` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8 refs/heads/main` ✓
+- `git diff --staged --stat` pre-commit = 3 files, +100/−3 (matches Stage 2 plan exactly — u1+u2 97L new + u3 4L + u4 2L = expected layout).
+- `git log -1 --stat` = `e60aacc`, kyeongmin, 3 files, +100/−3.
+- `git status --short -- src/` post-commit = empty (zero src/ delta confirmed).
+
+**commit_message_summary**: Conventional `docs(IMP-19):` prefix with `zone ratio reference + cross-link -- documented-axis close` headline. Body: per-file (new doc / backlog / insight-map) breakdown referencing A1-A5 axes and bidirectional cross-ref. Closing paragraph declares Stage 1/2 guardrail lock (src/ untouched, no role-based hardcoding into Phase Z, IMP-09 ownership preserved, soft-link integrity) and A5 dormancy gate. Co-Authored-By trailer.
+
+**guardrails (verified intact on commit)**:
+- GR1 no src/ edits: yes (`git status --short -- src/` empty pre + post commit).
+- GR2 docs-only references via file:line, no code body copying: yes (verified by visual inspection of IMP-19-ZONE-RATIO-REFERENCE.md).
+- GR3 no role-based `["배경","본심"]` hardcoding into Phase Z runtime: yes (pattern only quoted as read-only Phase O reference at A1).
+- GR4 soft-link integrity — IMP-19 status `documented`, IMP-09 carries back-ref, A5 gate documented: yes (backlog L53 + L70 confirm bidirectional ref).
+- GR5 IMP-09 boundary preserved — solver ownership not absorbed by IMP-19: yes (A4 statement explicit).
+- GR6 `build_containers_type_b` call sites untouched: yes (`git diff src/space_allocator.py` = empty).
+- workflow_atomicity_rules: 1 commit = 1 decision unit (= "IMP-19 documented carve-out"), same-file Edit sequential, scope-locked.
+- anchor_sync: not applicable (reference doc is new; PHASE-Q I4 row carries new anchor as forward reference, not SHA bump).
+- scope_lock: 3 files only; no scope bleed to working-tree concurrents.
+
+**unresolved_questions**: none. Stage 2 plan executed verbatim; Stage 4 PASS confirmed by Claude #N + Codex Stage 4 R1. IMP-09 activation status flip explicitly deferred (Stage 3 unresolved Q1 — outside IMP-19 scope).
+
+**evidence**:
+- `git log -1 --stat` confirms commit `e60aacc` author kyeongmin, 3 files, +100/−3.
+- `git ls-remote slide2 main` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8 refs/heads/main`.
+- `git ls-remote origin main` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8 refs/heads/main`.
+- `git push slide2 main` output: `02e2ae0..e60aacc  main -> main` (Gitea `C.E.L_Slide_test2`, `Processing 1 references` server ack).
+- `git push origin main` output: `02e2ae0..e60aacc  main -> main` (GitHub `keimin86/design_agent`).
+- `git diff --staged --stat` pre-commit = exactly 3 files, +100/−3, matching Stage 2 plan (u1+u2 = 97L new IMP-19 reference doc; u3 = 4L combined backlog row delta; u4 = 2L insight-map I4 row delta).
+- `git diff --staged --name-only` pre-commit = exactly the 3 paths above; no scope bleed.
+- runtime behavior change: 0 (no diff on src/renderer.py, src/space_allocator.py, src/phase_z2_pipeline.py, templates/**, tests/**; only docs/architecture/* touched).
+
+**source_comment_ids**: Claude #1 (Stage 1 R1); Claude #3 (Stage 2 R2); Claude #N (Stage 3 unit executor); Claude #N (Stage 4 test-verify R1 PASS); Codex Stage 4 R1 (PASS); Claude #5 (Stage 5 R1 — this comment).
+
+rewind_target: (n/a — push complete, both remotes verified)
+
+FINAL_CONSENSUS: YES
+
+=== EVIDENCE ===
+- commit_sha: `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8` (HEAD)
+- parent_sha: `02e2ae0` (Stage 1/2/3/4 entry)
+- files staged + committed (3, +100/−3): docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md (new, 97L, 5 axes A1-A5 + 7-row comparison table at A3); docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md (row L53 IMP-09 soft back-link added, row L70 IMP-19 status `pending`→`documented` + reference doc link); docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md (row L125 §3 I4 IMP-19 anchor + reference doc link prepended).
+- pre-commit `git diff --staged --stat` = exactly 3 files, +100/−3 (matches Stage 2 plan = u1+u2 97L new + u3 4L combined + u4 2L combined; A3 comparison table inflated total from estimated 74L → committed 97L per Stage 4 verifier acceptance).
+- pre-commit `git diff --staged --name-only` = exactly the 3 paths above; no scope bleed.
+- `git push slide2 main` = `02e2ae0..e60aacc  main -> main` (Gitea `Kyeongmin/C.E.L_Slide_test2` — issue #19 host, `Processing 1 references` server ack).
+- `git push origin main` = `02e2ae0..e60aacc  main -> main` (GitHub `keimin86/design_agent`).
+- post-push `git ls-remote slide2 main` = `e60aacc...` ✓ matches HEAD.
+- post-push `git ls-remote origin main` = `e60aacc...` ✓ matches HEAD.
+- runtime behavior change: 0 (no diff on src/renderer.py, src/space_allocator.py, src/phase_z2_pipeline.py, templates/**, tests/**; only docs/architecture/* touched).
+- guardrails verified: GR1 no src/ edits (`git status --short -- src/` empty), GR2 docs-only file:line refs (no code body copy), GR3 no role-based hardcoding into Phase Z runtime (Phase O `["배경","본심"]` only quoted as read-only reference at A1), GR4 soft-link integrity (bidirectional ref backlog L53 ↔ L70), GR5 IMP-09 solver ownership preserved (A4 boundary statement), GR6 `build_containers_type_b` call sites untouched.
+- excluded from staging (working tree untouched, separate axes): src/renderer.py, src/space_allocator.py, src/phase_z2_pipeline.py, orchestrator(원본).py, tests/PIPELINE.md, tests/PLAN.md, tests/PROGRESS.md, tests/README.md, tests/matching/*, chromedriver/win64/147.0.7727.117/, samples/uploads/, samples/mdx/05*.mdx, samples/mdx_batch/05.mdx, scripts/post_gitea_comment.py, scripts/sync_phase_z_run_to_frontend.py, templates/catalog/, templates/phase_z2/families/app_sw_package_vs_solution.html, templates/phase_z2/families/pre_construction_model_info_stacked.html, templates/phase_z2/frames/, .orchestrator/, .claude/settings.local.json.
+- scope_lock: 1 commit = 1 decision unit (IMP-19 documented carve-out), exactly 3 docs/architecture/* files; no working-tree concurrent bleed.

.orchestrator/drafts/19_stage_commit-push_codex_r1.md

@@ -0,0 +1,64 @@
+[Codex #6] Stage 5 commit-push verification R1
+
+Decision: PASS.
+
+I verified the Stage 5 commit/push claims against local Git state and both remotes. `HEAD` is `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`, with parent `02e2ae0afb7dab96f318de1fa18a661914101270`, and the commit subject is `docs(IMP-19): zone ratio reference + cross-link -- documented-axis close`.
+
+Remote verification:
+- `origin/main` resolves to `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`.
+- `slide2/main` resolves to `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`.
+- Local branch decoration also shows `HEAD -> main`, `origin/main`, and `slide2/main` all at `e60aacc`.
+
+Committed file scope:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` added, `+97`.
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` modified, `+1/-1`.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` modified, `+2/-2`.
+
+No unintended files were committed. `git diff-tree --no-commit-id --name-only -r HEAD -- src docs/architecture` returns only the three expected architecture docs. `git diff HEAD^ HEAD -- src/renderer.py src/space_allocator.py src/phase_z2_pipeline.py` returns no diff, so the Phase O/Q reference sources and Phase Z solver were not changed.
+
+Working tree note: the repository still has many pre-existing untracked local files under paths such as `.orchestrator/`, `chromedriver/`, `samples/`, `scripts/`, `templates/`, and `tests/matching/`. They are not part of `e60aacc`; this matches the Stage 5 exclusion claim.
+
+Content spot-check:
+- `IMP-19-ZONE-RATIO-REFERENCE.md` contains A1 consumer, A2 producer, A3 Phase Z solver delta, A4 IMP-09 boundary, and A5 re-activation gate / GR1-GR4.
+- `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` has the IMP-09 soft back-link to the IMP-19 reference doc and the IMP-19 row status is `documented`.
+- `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` has the IMP-19 I4 anchor and reference doc link; Step 8 / deterministic classification is preserved.
+
+No full pytest was run because the current stage task is commit/push verification and the prompt explicitly says not to run full pytest unless the stage requires it.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short`
+- `git rev-parse HEAD; git log -1 --pretty=format:'%H%n%h%n%P%n%s'`
+- `git show --stat --oneline --decorate --name-status HEAD`
+- `git remote -v`
+- `git ls-remote origin refs/heads/main`
+- `git ls-remote slide2 refs/heads/main`
+- `git show --numstat --format=fuller HEAD`
+- `git diff-tree --no-commit-id --name-only -r HEAD -- src docs/architecture`
+- `git diff HEAD^ HEAD -- src/renderer.py src/space_allocator.py src/phase_z2_pipeline.py`
+- `rg -n "^# IMP-19|A1|A2|A3|A4|A5|renderer\._group_blocks_by_area|build_containers_type_b|compute_zone_layout|IMP-09|re-activation|GR1|GR2|GR3|GR4" docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `rg -n "IMP-09|IMP-19|IMP-19-ZONE-RATIO-REFERENCE|documented|soft back-link" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `rg -n "IMP-19|I4|IMP-19-ZONE-RATIO-REFERENCE" docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `git status --short -- docs/architecture src`
+
+Files checked:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `src/renderer.py` (diff verification only)
+- `src/space_allocator.py` (diff verification only)
+- `src/phase_z2_pipeline.py` (diff verification only)
+
+Commit SHA:
+- `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+
+Verified facts:
+- `origin/main` equals `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`.
+- `slide2/main` equals `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`.
+- The commit contains exactly three files, all under `docs/architecture/`.
+- The commit has no diff for `src/renderer.py`, `src/space_allocator.py`, or `src/phase_z2_pipeline.py`.
+- The IMP-19 reference doc contains A1 through A5 and GR1 through GR4.
+- The backlog and insight-map cross-links are present.
+- Untracked local files remain outside the commit.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_final-close_claude_r1.md

@@ -0,0 +1,76 @@
+[Claude #4] Stage 6 final-close Round #1 — IMP-19 documented-axis close recommended.
+
+## 1. Issue body re-read (binding scope)
+- Step 8, INSIGHT-MAP §3 (Reference Only), priority ↓ low.
+- scope: `renderer._group_blocks_by_area` pattern reference + Phase Z zone-level ratio 분배 axis (별 axis 활성 시).
+- guardrail: Phase O 컨테이너 회귀 X · abstraction 다름 — 직접 통합 X.
+- dependency: soft link IMP-09.
+
+## 2. Commit verification on origin (and slide2 mirror)
+- `git rev-parse HEAD` → `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+- `git rev-parse origin/main` → `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+- `git rev-parse slide2/main` → `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+- All three aligned at the Stage 5 commit. No drift.
+
+Commit subject: `docs(IMP-19): zone ratio reference + cross-link -- documented-axis close`.
+
+`git diff-tree --no-commit-id --name-status -r HEAD`:
+- `A docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `M docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `M docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+
+`git diff HEAD^ HEAD -- src/` → empty. Runtime surface untouched (GR1 holds).
+
+## 3. Goal vs result
+| dimension | issue-body goal | Stage 5 result | match |
+|---|---|---|---|
+| scope axis | Reference-only pattern doc for `_group_blocks_by_area` + Phase Z zone-ratio axis hook | New `IMP-19-ZONE-RATIO-REFERENCE.md` (A1 consumer + A2 producer + A3 solver delta + A4 IMP-09 boundary + A5 re-activation gate) | ✓ |
+| guardrail — no Phase O regression | "직접 통합 X" | src/ diff empty; no role-hardcoding seeded | ✓ |
+| guardrail — abstraction separation | "abstraction 다름" | A3/A4 explicitly document block-level Type-B vs zone-level Phase Z delta | ✓ |
+| dependency — soft-link IMP-09 | Soft-link maintained | IMP-09 row in backlog carries back-link to IMP-19 doc; IMP-19 doc points to IMP-09 boundary in A4 | ✓ |
+| priority — ↓ low | Reference-only artifact, not solver work | Documented-axis (no runtime activation, A5 gates re-entry) | ✓ |
+
+## 4. Cross-ref / follow-up links integrity
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` — IMP-19 row status `documented` + reference doc link present (Stage 4 spot-check confirmed; remote `e60aacc` carries identical content per Codex Stage 5 R1).
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 I4 row (line 125) — IMP-19 anchor + ref doc link prepended; Step 8 + deterministic classification preserved.
+- `grep IMP-19 src/` → 0 matches (correct — docs-only axis).
+- `grep IMP-19 docs/architecture/` → 3 files: IMP-19 ref doc, backlog, insight-map. Bidirectional cross-ref intact (GR3 holds).
+
+## 5. Labels (current vs recommended)
+- Current: `needs-codex-review`.
+- Available repo labels: `blocked`, `closed`, `implemented`, `needs-claude-review`, `needs-codex-review`, `ready-for-implementation`, `scope-locked`, `verified`.
+- Precedent — IMP-18 (most recent documented-axis close) closed with `needs-codex-review` label intact, no `implemented`/`verified` added (closed_at 2026-05-19T09:45:08+09:00). I follow the same convention: do not mutate labels in this round; close action is sufficient.
+- If repo policy later wants `implemented` + `verified` on documented-axis closes, propose a sweep PR alongside IMP-18 — out of scope here.
+
+## 6. Close decision
+**Recommend CLOSE** as documented-axis. Rationale:
+- Stage 4 PASS (3-axis test: scope_verification, src_isolation, unit u1~u4 spot-check, source-anchor spot-check — all pass; regression_check none).
+- Stage 5 PASS (commit `e60aacc` pushed to `origin/main` + `slide2/main`; src/ diff empty).
+- Goal (reference-only artifact + cross-ref + IMP-09 boundary statement) fully delivered.
+- All guardrails GR1-GR4 hold.
+- A5 re-activation gate documents the only re-open path (IMP-09 cannot handle zone-level ratio OR Type-B legacy returns to runtime).
+
+The actual `state: closed` flip is left for the orchestrator after Codex verifier YES on this round, per consensus protocol.
+
+## 7. Out of scope (do not address in this round)
+- IMP-09 status row state change (`pending` → active) — Stage 3 unresolved_question, deferred to IMP-09 owner.
+- Any retroactive label policy normalization across closed docs-only issues.
+- Phase Z solver runtime change (IMP-09 territory).
+
+=== EVIDENCE ===
+- `git rev-parse HEAD` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+- `git rev-parse origin/main` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+- `git rev-parse slide2/main` = `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+- `git log -1 --format='%H%n%s%n%b' HEAD` confirms subject `docs(IMP-19): zone ratio reference + cross-link -- documented-axis close` + body lists 3 docs only.
+- `git diff-tree --no-commit-id --name-status -r HEAD` = A IMP-19-ZONE-RATIO-REFERENCE.md, M PHASE-Q-INSIGHT-TO-22STEP-MAP.md, M PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md.
+- `git diff HEAD^ HEAD -- src/` = empty.
+- `wc -l docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` = 97 lines.
+- `grep -r IMP-19 src/` = 0 matches.
+- `grep -r IMP-19 docs/architecture/` = 3 files (ref doc + backlog + insight-map).
+- GET `/api/v1/repos/Kyeongmin/C.E.L_Slide_test2/issues/19` → state=open, labels=['needs-codex-review'].
+- GET `/api/v1/repos/Kyeongmin/C.E.L_Slide_test2/issues/18` → state=closed, labels=['needs-codex-review'] (documented-axis precedent).
+- Stage 4 exit report (PASS) + Stage 5 exit report (commit_sha=e60aacc, push PASS) read as binding contracts.
+
+source_comment_ids: Claude #4 (Stage 6 final-close R1); upstream Codex #6 (Stage 5 push verification R1 PASS); Codex Stage 4 R1 (test-verify PASS).
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_final-close_codex_r1.md

@@ -0,0 +1,60 @@
+[Codex #N] Stage 6 final-close Round #1 independent verification.
+
+Decision: PASS. Issue #19 can close as a documented-axis reference task.
+
+Scope check:
+- HEAD, origin/main, and slide2/main are aligned at e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8.
+- The committed change set is exactly three architecture docs: one new IMP-19 reference doc plus backlog and insight-map cross-reference edits.
+- `git diff HEAD^ HEAD -- src/` is empty, so runtime code was not changed.
+- `rg -n "IMP-19|IMP-19-ZONE-RATIO-REFERENCE" src/` returns no matches, so IMP-19 identifiers did not enter runtime code.
+
+Content check:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` contains A1 consumer trace for `renderer._group_blocks_by_area`, A2 producer trace for `space_allocator.build_containers_type_b`, A3 Phase Z solver delta, A4 IMP-09 boundary, and A5 re-activation gate with GR1-GR4.
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` has the IMP-19 row marked `documented` with the reference doc link, and the IMP-09 row has a soft back-link to the IMP-19 reference doc.
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` has the I4 row anchored as `IMP-19 I4 zone 비중 분배` with the reference doc link, while preserving Step 8 and deterministic classification.
+- Source anchors cited by the doc resolve to live code: `src/renderer.py:210`, `:234`, `:267`, `:353`, `:426`; `src/space_allocator.py:445`; `src/phase_z2_pipeline.py:794`, `:924`, `:1094`, `:1125`, `:1155`, `:1210`, `:1452`.
+
+Guardrail check:
+- GR1 holds: no `src/` edits in the pushed commit.
+- GR2 holds: IMP-19 is docs-only and references file:line anchors rather than copying code bodies.
+- GR3 holds: cross-links exist in the IMP-19 reference doc, the backlog, and the insight-map I4 row.
+- GR4 holds: IMP-19 remains a dormant documented reference, with solver ownership left to IMP-09 and re-activation gated by A5.
+
+=== EVIDENCE ===
+Commands run:
+- `git rev-parse HEAD`
+- `git rev-parse origin/main`
+- `git rev-parse slide2/main`
+- `git diff-tree --no-commit-id --name-status -r HEAD`
+- `git diff HEAD^ HEAD -- src/`
+- `rg -n "^# IMP-19|A1|A2|A3|A4|A5|_group_blocks_by_area|build_containers_type_b|compute_zone_layout|compute_zone_layout_cols|re-activation|GR1|GR2|GR3|GR4|IMP-09" docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `rg -n "IMP-19|IMP-09|documented|IMP-19-ZONE-RATIO-REFERENCE" docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `rg -n "I4|IMP-19|IMP-19-ZONE-RATIO-REFERENCE|Step 8|deterministic" docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `rg -n "IMP-19|IMP-19-ZONE-RATIO-REFERENCE" src/`
+- `rg -n "def _group_blocks_by_area|role_order|height_px|_group_blocks_by_area\(|build_containers_type_b|def compute_zone_layout\(|def compute_zone_layout_cols\(|_build_rows_dynamic|_build_cols_dynamic|_build_grid_dynamic_2d|_build_fr_default" src/renderer.py src/space_allocator.py src/phase_z2_pipeline.py`
+- `git diff-tree --no-commit-id --name-only -r HEAD -- src docs/architecture`
+- `git status --short`
+- `git show --stat --oneline --name-status HEAD`
+
+Files checked:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `src/renderer.py`
+- `src/space_allocator.py`
+- `src/phase_z2_pipeline.py`
+
+Commit SHA:
+- `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`
+
+Verified facts:
+- Local HEAD, origin/main, and slide2/main all point to `e60aacc3dc5c2ecdb435b85e48c2264ffeeaaeb8`.
+- The commit changes only `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`, `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`, and `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`.
+- `git diff HEAD^ HEAD -- src/` is empty.
+- IMP-19-specific identifiers have zero matches under `src/`.
+- The new reference doc includes A1, A2, A3, A4, A5, and GR1-GR4.
+- The backlog has an IMP-19 documented row and an IMP-09 soft back-link to the IMP-19 reference doc.
+- The insight-map I4 row includes the IMP-19 anchor and reference doc link while preserving Step 8 and deterministic classification.
+- Existing untracked local paths are outside the pushed commit and do not affect this final-close decision.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_problem-review_claude_r1.md

@@ -0,0 +1,139 @@
+[Claude #1] Stage 1 — problem-review Round #1
+
+## 1. Re-read of scope (backlog §2 IMP-19)
+
+source = INSIGHT-MAP §3 / §2.8 I4 (`renderer._group_blocks_by_area`) — Reference Only axis.
+related step = Step 8 (Zone + Region Ratio Planning).
+priority = ↓ low. dependency = `soft link: IMP-09` (zone 비중 분배 영역 공유).
+guardrail = Phase O 컨테이너 회귀 X / 직접 통합 X.
+
+## 2. Root cause (3-axis check)
+
+### Axis A — Phase O / Phase Q source pattern (read-only reference)
+
+- `D:\ad-hoc\kei\design_agent\src\renderer.py:210-295` `_group_blocks_by_area(blocks, container_specs)` :
+  - block 들을 `block["area"]` 키로 OrderedDict grouping ("body", "left", "right", "hero", "detail" 등).
+  - `container_specs` 가 들어오면 그 area zone 안에 `role_order = ["배경", "본심"]` 순서로 *역할별 고정 높이 컨테이너* 생성 (`spec.zone` / `spec.topic_ids` 매칭, fallback = 순서 매칭).
+  - 각 container 의 inline style = `height:{spec.height_px}px` + `font-size:{font_size}px` + `--spacing-inner:{padding}px` + `--font-body:{font_size/16:.3f}rem` (block_constraints 에서 가져옴).
+  - 미배정 block 은 container 옆에 그대로 append.
+  - `len(block_list) == 1` → single html. 그 외 → flex-column wrapper.
+- 호출 지점 = `renderer.py:353` `render_multi_page()` + `renderer.py:426` `render_slide()` (모두 Phase O / Phase Q 흐름).
+- container_specs 의 생산자 = `D:\ad-hoc\kei\design_agent\src\space_allocator.py:445` `build_containers_type_b(page_structure, ...)` — Kei `page_structure` 판단 (역할 → zone + topic_ids + weight) 에서 *역할별 ContainerSpec* 생성. `top` / `bottom_left` / `bottom_right` / `footer` zone semantics (Phase Q 유형 B).
+
+### Axis B — Phase Z Step 8 zone-ratio 현 상태 (active, IMP-09 cover)
+
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py:794-853` `compute_zone_layout(zones_data, ...)` :
+  - row-axis solver — `min_height_first + content_weight_distribution`.
+  - frame contract `visual_hints.min_height_px` 우선 + 남은 공간을 `content_weight.score` 비율로 분배 + rounding diff 마지막 zone 흡수.
+- `phase_z2_pipeline.py:924-` `compute_zone_layout_cols` — col-axis 대칭 helper (IMP-09 PR 1).
+- `phase_z2_pipeline.py:1125-1152` `_build_rows_dynamic` (topology="rows", horizontal-2), `1210-` `_build_cols_dynamic` (topology="cols", vertical-2), `1155-1207` `_build_grid_dynamic_2d` (T / inverted-T / side-T-left / side-T-right / 2x2 — IMP-09 PR 2).
+- `phase_z2_pipeline.py:1343-1452` `build_layout_css(layout_preset, zones_data, override_zone_geometries)` — topology dispatch + user override (`_override_to_grid_tracks`).
+- 결론 : Phase Z Step 8 의 *normal-path zone-비중 분배* = `compute_zone_layout` + `compute_zone_layout_cols` + 8-preset dispatcher 로 완비. IMP-09 PR 1 + PR 2 이미 implemented (`feat(IMP-09): PR 1 — col-axis solver + per-zone geometry mapper + retry gate` / `feat(IMP-09): PR 2 — 2-D dynamic dispatch for 5 preset families`).
+
+### Axis C — abstraction gap (audit §2.8 I4 명시)
+
+| Phase Q (`_group_blocks_by_area`) | Phase Z (`compute_zone_layout` + `build_layout_css`) |
+|---|---|
+| 단위 = block-level area (body / left / right / hero / detail) | 단위 = zone position (preset 의 positions 키 — top / bottom_left / bottom_right 등) |
+| 비중 source = Kei `page_structure` → `build_containers_type_b` → `ContainerSpec.height_px` (역할별 px 직접 지정) | 비중 source = frame contract `min_height_px` + content_object 의 `content_weight.score` (deterministic) |
+| 의미 단위 = 역할 ("배경" / "본심") — Kei semantics | 의미 단위 = layout preset positions + content-weight (preset 어휘) |
+| AI 의존 = Kei `page_structure` (Phase Q AI 호출 path) | AI 의존 = 0 (normal path AI=0 invariant PZ-1) |
+| catalog = `templates/blocks/catalog.yaml` (Phase Q block id → variant) | catalog = `templates/phase_z2/catalog/frame_contracts.yaml` (frame contract + frame_partials) |
+
+audit §2.8 I4 (`PHASE-Q-AUDIT.md` L858) : *"_group_blocks_by_area (area 비중 grouping + Phase O 컨테이너) 가 Phase Z zone 비중 분배 axis 활성 시 Reference Only. abstraction 다름 — 직접 통합 X."*
+
+**Root cause framing** : IMP-19 = Phase Q `_group_blocks_by_area` 의 *area + 역할별 컨테이너 height grouping* 패턴을 Phase Z zone-비중 분배 axis 가 활성될 때 *reference 만* 하라는 dormant marker. Phase Z 의 Step 8 normal-path zone-비중 분배는 IMP-09 PR 1 / PR 2 가 이미 cover (8-preset deterministic dispatcher + min_height + content_weight 분배). 따라서 IMP-19 의 *active implementation target = 부재*. 활성 trigger = (가설) Phase Z 에 *역할 (semantic role) 기반 추가 컨테이너 layer* 가 zone 안에 도입되어 Phase O 의 "배경 / 본심" 같은 sub-zone height 분배가 필요해질 때 — 현재 없음.
+
+## 3. Scope-lock proposal
+
+### (a) 본 issue 의 산출물 = **doc-only Reference Only carve-out** (IMP-18 mirror)
+
+direct code change 없음. 산출물 위치 후보 (Stage 2 결정 대상):
+
+- Option A — 신설 doc : `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md` (IMP-18 `IMP-18-SVG-GAP-REPORT.md` 패턴 mirror — 별 sub-axis 별 reference doc).
+- Option B — 기존 doc entry 보강 : `docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` Step 8 row + `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` §2 IMP-19 row status 갱신 + audit §2.8 I4 cross-ref.
+
+Stage 2 plan 단계에서 lock. 현 turn 선호 = Option A (IMP-17 / IMP-18 mirror — Reference Only / dormant axis 의 doc 분리 가독성 ↑, backlog row 와 1:1 link).
+
+### (b) 본 issue 의 산출물 내용 (확정 boundary)
+
+1. **Pattern source path 명시** (현재 코드 기준 — Stage 2 fact-verify 대상) :
+   - `src/renderer.py:210-295` (`_group_blocks_by_area`, 86 lines, `container_specs` 적용 분기 + role_order = `["배경", "본심"]` + topic_id 매칭 fallback).
+   - `src/renderer.py:353`, `src/renderer.py:426` (호출 지점 — `render_multi_page` / `render_slide`).
+   - `src/space_allocator.py:445+` (`build_containers_type_b` — `page_structure` → ContainerSpec 의 생산자, Kei AI 의존 path).
+2. **Data contract 기술** (input/output) :
+   - input = `blocks: list[{area, html, _topic_id}]` + `container_specs: dict[role_name → ContainerSpec(zone, topic_ids, height_px, block_constraints)]`.
+   - output = `list[{area, html}]` — area-grouped, container-wrapped HTML.
+3. **Phase Z 적용 boundary (G3 = 직접 통합 X)** :
+   - Phase Z code 에서 `from src.renderer import _group_blocks_by_area` **직접 import 금지**.
+   - Phase Z code 에서 `from src.space_allocator import build_containers_type_b` **직접 import 금지** (Kei `page_structure` 의존).
+   - 활성 시점에 Phase Z 자체 helper (가설 module: `phase_z2_pipeline` 내부 helper) 로 *abstraction 변환된 mirror* 가 필요 — 직접 import = Phase Q regression risk (audit §0-A invariant + Phase O 컨테이너 회귀 guardrail).
+   - Phase Z 의 자체 consumer 후보 = (가설) Step 8 zone-level dispatcher 가 zone 내부에 sub-zone semantic role layer 를 도입해야 할 axis — 현재 axis 부재 (IMP-09 가 cover).
+4. **Activation trigger 명시** :
+   - Phase Z 에 *zone 내부 sub-zone semantic role layer* 가 필요한 새 axis 가 lock 되는 시점 (예: 가설 "Phase Z 안에서도 '배경 / 본심' 같은 역할별 height share 가 필요한 frame 등록" — 현재 catalog 의 frame contract 들은 sub_zones layer 가 *position 기반* 으로만 동작, semantic role layer 없음).
+   - IMP-09 가 cover 하지 않는 case 가 새 issue 로 발생할 때 본 reference doc 이 patten reference 로 link 됨.
+   - 본 reference doc 박힘 자체로 IMP-19 implementation = 종료.
+
+### (c) IMP-09 ↔ IMP-19 boundary (soft link 해석)
+
+- IMP-09 (B-4 다른 layout zone-geometry) = `build_layout_css` 분기 확장 (8-preset deterministic dispatcher). **implemented** (PR 1 / PR 2 commit log 확인).
+- IMP-19 (I4 zone 비중 분배) = `_group_blocks_by_area` Reference Only marker. **abstraction 다름** — IMP-09 와 *영역 공유* (zone-비중 분배) 하지만 *abstraction 분리* (Phase Q area + 역할 vs Phase Z preset + content_weight). IMP-09 의 산출물에 IMP-19 의 Phase O role-layer 를 *후행 합산* 할 시점이 미정 → 현 시점 IMP-19 = dormant.
+
+## 4. Guardrails (issue body 명시 + 추가 lock)
+
+| # | guardrail | 근거 |
+|---|---|---|
+| G1 | `src/renderer.py` 코드 수정 X (특히 `_group_blocks_by_area` 본체) | issue body 명시 — Phase O 컨테이너 회귀 X |
+| G2 | `src/space_allocator.py` 코드 수정 X (특히 `build_containers_type_b`) | Phase Q 유형 B ContainerSpec 생산자, pattern 보존 |
+| G3 | Phase Z code (`src/phase_z2_*.py`) 에서 `from src.renderer import _group_blocks_by_area` 직접 import 금지 / `from src.space_allocator import build_containers_type_b` 직접 import 금지 | audit §2.8 I4 명시 "직접 통합 X" + Phase R'/Phase O 격리 |
+| G4 | normal path AI 호출 추가 X | PZ-1 invariant (AI=0 normal path) — Phase Q `page_structure` Kei AI 의존 패턴 mirror 금지 |
+| G5 | `templates/phase_z2/catalog/frame_contracts.yaml` 신규 entry 추가 X | IMP-04 axis 영역, 본 issue scope 밖 |
+| G6 | IMP-09 implemented 결과 (`compute_zone_layout` / `_build_*` / `build_layout_css`) 코드 수정 X | IMP-09 axis 영역, 본 issue scope 밖 |
+| G7 | doc 산출물에 hardcoded MDX 03/04/05 sample 의존 X | RULE 0 PIPELINE-CONSTRUCTION (general pattern reference 만) |
+| G8 | doc 안에 Kei `page_structure` semantics ("배경 / 본심") 를 Phase Z 권장 패턴으로 박지 말 것 | PZ §0 invariant — Kei composition / Kei slot fill 회귀 금지 |
+
+## 5. Verification axis (Stage 2/3 가 lock 해야 할 것)
+
+- Stage 2 plan 단계에서 :
+  - doc 위치 lock (Option A vs Option B).
+  - `src/renderer.py:210-295` 의 현재 line 범위 fact-verify (PR drift 가능).
+  - `src/space_allocator.py:445` 의 `build_containers_type_b` 시그니처 fact-verify.
+  - IMP-09 implemented 결과 cross-ref link path (`compute_zone_layout` / `_build_rows_dynamic` / `_build_cols_dynamic` / `_build_grid_dynamic_2d` line number 확인).
+  - audit §2.8 I4 line / INSIGHT-MAP §3 row 125 / backlog §2 IMP-19 row anchor 정합.
+- Stage 3 edit 단계 :
+  - doc 산출물 1 개 박기 (Option A 시 신규 file, Option B 시 기존 doc 보강).
+  - backlog §2 IMP-19 row status `pending → documented` (IMP-18 와 같은 pattern — Reference Only 산출물 박힘 시점에 status 갱신).
+- Stage 4 verify :
+  - pattern source path/line 의 *현재 코드 일치* (`renderer.py:210` 존재 / `space_allocator.py:445` 존재) 재확인.
+  - doc cross-ref link 깨짐 없음 (`PHASE-Q-AUDIT.md` §2.8 I4 L858 / `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 line 125 / `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` §2 IMP-19 row).
+  - normal path code unchanged (`git diff src/` empty for `phase_z2_*` 와 `renderer.py` / `space_allocator.py`).
+  - `pytest -q` regression 0 (doc-only carve-out — test 회귀 없어야 정상).
+
+## 6. Open question (Codex round 1차 review 대상)
+
+- **Q1** — doc 위치 : Option A (신규 `IMP-19-ZONE-RATIO-REFERENCE.md`) vs Option B (기존 doc entry 보강) — IMP-17 / IMP-18 패턴 mirror 시 Option A 가 backlog 운영 컨벤션상 권장인지 확인.
+- **Q2** — IMP-09 ↔ IMP-19 boundary 가 audit §2.8 I4 의 "활성 시 Reference Only" wording 만으로 충분히 명확한지, doc 산출물에서 "현 시점 IMP-19 active 구현 target = 0 (IMP-09 가 cover)" 을 명시적으로 박는 게 적절한지.
+- **Q3** — `_group_blocks_by_area` 의 role-layer ("배경 / 본심") semantics 가 Kei `page_structure` 에서 옴 → 본 reference doc 에서 *Phase Z 가 이 semantics 를 채택할 자유는 없다 (PZ §0 Kei composition 금지)* 를 G8 으로 박은 게 과/부족한지 — Phase Z 의 *순수 geometric role layer* (예: visual-emphasis tier) 가 미래에 등장할 가능성을 doc 에 어디까지 열어둘지.
+
+=== EVIDENCE ===
+
+- Files read :
+  - `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (§2 IMP-19 row L70 — soft link IMP-09, status pending).
+  - `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (§3 line 125 — `I4 zone 비중 분배 | Step 8 | §2.8 I4 ...`; §2 line 84 Step 8 row).
+  - `docs/architecture/PHASE-Q-AUDIT.md` (§2.8 I4 L796 / L818 / L832 / L858 — `_group_blocks_by_area` 분류 + abstraction 다름 + 직접 통합 X).
+  - `docs/architecture/IMP-18-SVG-GAP-REPORT.md` (전체 — Reference Only carve-out doc pattern reference).
+  - `src/renderer.py` (L210-295 `_group_blocks_by_area`, L353 / L426 호출 지점).
+  - `src/space_allocator.py` (L445+ `build_containers_type_b` 시그니처 + page_structure 소비).
+  - `src/phase_z2_pipeline.py` (L794-853 `compute_zone_layout`, L1125-1152 `_build_rows_dynamic`, L1155-1207 `_build_grid_dynamic_2d`, L1210+ `_build_cols_dynamic`, L1343-1452 `build_layout_css`).
+- Git log evidence (IMP-09 implemented 확인) :
+  - `1fb9732 feat(IMP-09): PR 2 — 2-D dynamic dispatch for 5 preset families`
+  - `201099e feat(IMP-09): PR 1 — col-axis solver + per-zone geometry mapper + retry gate`
+- Grep results :
+  - `_group_blocks_by_area` = `src/renderer.py:210, 353, 426` 만. Phase Z 모듈 (`src/phase_z2_*.py`) 0 매치.
+  - `build_containers_type_b` = `src/space_allocator.py:445` 정의 + `src/pipeline.py:531-532` 단일 호출 (Phase Q 흐름) 만.
+- `.orchestrator/issues/19_comments_cache.json` = `[]` — 본 라운드 = 첫 라운드 (Round 1).
+- No code change 발생 — Stage 1 problem-review.
+
+rewind_target: 해당 없음 (YES 제안)
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_problem-review_codex_r1.md

@@ -0,0 +1,72 @@
+[Codex #1] Stage 1 problem-review Round #1
+
+## Scope read
+
+Stage task is root-cause + scope-lock verification only. I used local grep/range reads and did not run pytest.
+
+## 3-axis verification
+
+Axis A - Phase O / Phase Q reference pattern:
+
+- Confirmed `src/renderer.py:210-295` defines `_group_blocks_by_area(blocks, container_specs=None)`.
+- It groups blocks by `block["area"]` in insertion order, then, only when `container_specs` exists and area is one of `body/left/right/hero/detail`, creates fixed-height role containers from `ContainerSpec`.
+- The role loop is hardcoded to `["배경", "본심"]`; blocks are matched by `_topic_id in spec.topic_ids`, then by fallback order, with unassigned blocks appended raw.
+- Container inline style uses `height:{spec.height_px}px`, `font-size` and spacing variables from `spec.block_constraints`; this is block/role-level container composition, not Phase Z zone geometry.
+- Confirmed call sites: `render_multi_page()` passes `_container_specs` at `src/renderer.py:352-353`; `render_slide()` calls `_group_blocks_by_area(blocks_raw)` without container specs at `src/renderer.py:426`.
+- Confirmed upstream Type B container construction at `src/space_allocator.py:445-586`: `build_containers_type_b(page_structure, ...)` consumes role `zone/topic_ids/weight`, derives top/bottom/footer heights, and returns role keyed `ContainerSpec`s. This is a reference pattern only under IMP-19, not a direct integration target.
+
+Axis B - Phase Z Step 8 active zone-ratio path:
+
+- Confirmed `src/phase_z2_pipeline.py:794-853` has the row-axis solver `compute_zone_layout`: min-height first, remaining height distributed by `content_weight.score`, rounding diff absorbed by the last zone, and trace fields returned.
+- Confirmed `src/phase_z2_pipeline.py:924-972` has `compute_zone_layout_cols`: column widths distributed by `content_weight.score`, with zero-weight equal fallback and trace fields.
+- Confirmed `src/phase_z2_pipeline.py:1125-1452` dispatches the active CSS layout path: rows topology -> `_build_rows_dynamic`, cols topology -> `_build_cols_dynamic`, and `T/inverted-T/side-T-left/side-T-right/2x2` -> `_build_grid_dynamic_2d`; geometry overrides for 2-D presets go through `_override_to_grid_tracks`.
+- Confirmed `templates/phase_z2/layouts/layouts.yaml:30-130` defines the 8 preset vocabulary: `single`, `horizontal-2`, `vertical-2`, `top-1-bottom-2`, `top-2-bottom-1`, `left-1-right-2`, `left-2-right-1`, `grid-2x2`.
+- Important caveat: `single` still falls to `_build_fr_default` (`src/phase_z2_pipeline.py:1094-1122`, `1451-1452`), which is correct as there is no competing zone-ratio distribution for one zone.
+- Important caveat: `docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md:243/389/413` still says only `horizontal-2` is dynamic. That doc statement is stale relative to the current code and the IMP-09 implementation claims; it should not be used to reopen IMP-19 as a normal-path zone-layout implementation task.
+
+Axis C - remaining abstraction gap:
+
+- Verified the backlog row says IMP-19 is Reference Only and pending: `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:70` maps `renderer._group_blocks_by_area` to a zone-level ratio distribution reference, with guardrails `Phase O container regression X / direct integration X` and soft link to IMP-09.
+- Verified the insight map also labels I4 as Step 8, pending, deterministic: `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:84` and `:125`.
+- Verified runtime Step 8 still records placeholder region-level signals: `src/phase_z2_pipeline.py:3982-4080` sets `region_count=1`, emits passive region/display candidates, and explicitly marks region-level ratio as partial/equal placeholder dependent on Step 3/4.
+- Verified `src/phase_z2_internal_region_planner.py:1-20` says the B2 planner is dormant and not connected to runtime; `src/phase_z2_placement_planner.py:1-12` says the B4 placement planner is also dormant. So region-level ratio planning is a separate future axis, not solved by directly porting `_group_blocks_by_area`.
+
+## Scope lock
+
+Root cause is not a missing direct call from Phase Z into `renderer._group_blocks_by_area`. The direct integration path is explicitly forbidden by the issue guardrail and would couple Phase Z zone geometry to Phase O role containers.
+
+The current Stage 1 scope should be locked as:
+
+- Accept IMP-09 as covering the active Phase Z normal-path zone-level ratio distribution across the multi-zone presets, with `single` using the fr-default one-zone sink.
+- Keep IMP-19 as a low-priority reference/abstraction audit item: compare the useful pattern from `_group_blocks_by_area` (input grouping -> ratio/fixed-size container plan -> traceable composed output) against Phase Z Step 8 contracts, without reusing Phase O container code.
+- Do not modify Phase O containers, do not directly integrate `ContainerSpec`/`build_containers_type_b` into Phase Z, and do not hardcode MDX 03/04/05 behavior.
+- If this proceeds to Stage 2, the plan should be doc/contract oriented unless the team explicitly chooses to activate the separate region-level axis. Any implementation plan must name the stale overview lines above and either update them or explain why docs are out of scope.
+
+No missing source files were found for the referenced symbols. The only wrong/stale assumption found is the architecture overview's outdated "horizontal-2 only dynamic" wording.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\.orchestrator\tmp\codex_prompt.txt'`
+- `rg -n "def _group_blocks_by_area|_group_blocks_by_area\(|def render_multi_page|def render_slide|container_specs|ContainerSpec" src\renderer.py src\space_allocator.py`
+- `rg -n "def compute_zone_layout|def compute_zone_layout_cols|def _build_rows_dynamic|def _build_grid_dynamic_2d|def _build_cols_dynamic|def build_layout_css|_override_to_grid_tracks|topology" src\phase_z2_pipeline.py`
+- `rg -n "IMP-19|I4|_group_blocks_by_area|Step 8|Zone \+ Region|zone.*ratio|ratio" docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md`
+- `rg -n "LAYOUT_PRESETS|horizontal-2|vertical-2|top-1-bottom-2|left-1-right-2|grid-2x2|topology" src templates\phase_z2 -g "*.py" -g "*.yaml" -g "*.yml"`
+- `rg -n "internal_regions|region_level|region-level|ratio_estimate|Internal Region|region_ratios|region_count" src\phase_z2_pipeline.py src\phase_z2_composition.py src templates\phase_z2 docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md`
+- `rg -n "_container_specs|build_containers_type_b|calculate_container_specs|page_structure|container_specs" src -g "*.py"`
+- Range reads of `src\renderer.py`, `src\space_allocator.py`, `src\phase_z2_pipeline.py`, `templates\phase_z2\layouts\layouts.yaml`, `src\phase_z2_internal_region_planner.py`, and `src\phase_z2_placement_planner.py`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\src\renderer.py`
+- `D:\ad-hoc\kei\design_agent\src\space_allocator.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_composition.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_internal_region_planner.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_placement_planner.py`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\layouts\layouts.yaml`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md`
+
+Test results: not run; Stage 1 explicitly says NO pytest.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_simulation-plan_claude_r1.md

@@ -0,0 +1,56 @@
+[Claude #2] Stage 2 — Simulation Plan R1 (compact)
+
+=== IMPLEMENTATION_UNITS ===
+- id: u1
+  summary: Create docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md — header + A1 (renderer._group_blocks_by_area consumer at src/renderer.py:210-295, call sites :352-353 and :426) + A2 (space_allocator.build_containers_type_b producer trace at src/space_allocator.py:445-586). Read-only file:line refs only.
+  files:
+    - docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md
+  tests: []
+  estimate_lines: 38
+- id: u2
+  summary: Append A3 (Phase Z solver delta vs Phase O block-level container — src/phase_z2_pipeline.py:794-853 row + :924-972 col + :1125-1452 dispatch, IMP-09 owned) + A4 (IMP-09 boundary statement) + A5 (activation gate / re-activation triggers) to IMP-19-ZONE-RATIO-REFERENCE.md.
+  files:
+    - docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md
+  tests: []
+  estimate_lines: 36
+- id: u3
+  summary: Update docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md — IMP-19 row (line 70) append doc link + status pending → documented; IMP-09 row (line 53) append back-reference to IMP-19 doc.
+  files:
+    - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+  tests: []
+  estimate_lines: 4
+- id: u4
+  summary: Update docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md line 125 — prepend IMP-19 ID anchor + reference doc link into existing §3 I4 row. No other row content change.
+  files:
+    - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+  tests: []
+  estimate_lines: 2
+
+Per-unit rationale
+- u1: Doc skeleton + two source pattern sections (consumer A1, producer A2). Mirrors IMP-18-SVG-GAP-REPORT.md precedent. No src/ edits.
+- u2: Append-only — A3 surfaces block-level vs zone-level distinction; A4 upholds IMP-09 ownership; A5 documents re-activation preconditions.
+- u3: Bidirectional cross-ref + status. `documented` follows IMP-18 backlog precedent (line 69). IMP-09 row gets back-ref only.
+- u4: ID anchor parity with IMP-17 at INSIGHT-MAP line 123. Pure prefix + link.
+
+Out-of-scope (Stage-1 lock GR1-GR6)
+- Any modification to src/renderer.py, src/space_allocator.py, src/phase_z2_pipeline.py.
+- Introducing role-based ["배경","본심"] hardcoding into Phase Z; replacing min_height_first + content_weight solver; migrating ContainerSpec.block_constraints; changing build_containers_type_b call sites.
+- Tests — Stage 1 precludes code change; zero runtime surface.
+
+Rollback
+1 unit = 1 commit. u1/u2 → `git rm` of the new doc. u3/u4 → `git checkout HEAD -- <file>`.
+
+=== EVIDENCE ===
+- src/renderer.py:210-295 consumer verified (Read): OrderedDict L218-223, role_order=["배경","본심"] L234, topic_id+fallback match L240-253, inline-style injection L265-273, len==1 branch L283-291.
+- src/renderer.py:352-353 and :426 call sites — Stage 1 binding contract.
+- src/space_allocator.py:445-586 producer verified (Read): page_structure→ContainerSpec; zone classification top/bottom/bottom_left/bottom_right/footer at L478-484; slide-base geometry top:65 h:590 at L489-490.
+- src/phase_z2_pipeline.py:794-853 row solver verified (Read): min_height_first L807-819, content_weight L823-826, rounding L828-832, trace dict L836-852.
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:70 IMP-19 row carries `soft link: IMP-09` + `pending` (Grep -n).
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:53 IMP-09 row carries `none` dep + `implemented` (Grep -n).
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md:125 §3 I4 row currently lacks IMP-19 prefix (Grep -n).
+- IMP-18-SVG-GAP-REPORT.md = 64 lines, A1/A2/A3 structural precedent (wc -l + Read).
+- IMP-17-CARVE-OUT.md = 44 lines, naming-convention precedent (wc -l).
+- Stage 1 contract: scope_lock 1-3, out_of_scope 1-5, GR1-GR6 respected.
+- Format: no inline code; body well under 5,000 chars; rationale ≤ 3 lines per unit.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_simulation-plan_claude_r2.md

@@ -0,0 +1,59 @@
+[Claude #3] Stage 2 — Simulation Plan R2 (compact, format-fixed)
+
+Format fix vs R1: removed pre-units analysis sections; no fenced code blocks; complete evidence block.
+
+=== IMPLEMENTATION_UNITS ===
+- id: u1
+  summary: Create docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md — header + A1 (renderer._group_blocks_by_area consumer at src/renderer.py:210-295, call sites :352-353 with container_specs and :426 without) + A2 (space_allocator.build_containers_type_b producer trace at src/space_allocator.py:445-586). Read-only file:line refs; no code copying.
+  files:
+    - docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md
+  tests: []
+  estimate_lines: 38
+- id: u2
+  summary: Append A3 (Phase Z solver delta vs Phase O block-level container — src/phase_z2_pipeline.py:794-853 row + :924-972 col + :1125-1452 dispatch, IMP-09 owned) + A4 (IMP-09 boundary statement) + A5 (re-activation gate + GR1-GR4) to IMP-19-ZONE-RATIO-REFERENCE.md.
+  files:
+    - docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md
+  tests: []
+  estimate_lines: 36
+- id: u3
+  summary: Update docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md — IMP-19 row append doc link + status pending → documented; IMP-09 row append back-reference to IMP-19 doc.
+  files:
+    - docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+  tests: []
+  estimate_lines: 4
+- id: u4
+  summary: Update docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md §3 I4 row (line 125) — prepend IMP-19 ID anchor + reference doc link. No other row content change.
+  files:
+    - docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+  tests: []
+  estimate_lines: 2
+
+Per-unit rationale
+- u1: Doc skeleton + consumer A1 + producer A2. Mirrors IMP-18-SVG-GAP-REPORT.md precedent (standalone doc under docs/architecture/). No src/ edits per GR1/GR2.
+- u2: Append-only — A3 surfaces block-level vs zone-level abstraction; A4 names IMP-09 as solver owner; A5 documents when IMP-19 may re-open (GR4 soft-link integrity).
+- u3: Bidirectional cross-ref + status. `documented` follows IMP-18 backlog precedent. IMP-09 row gets back-ref only (no status change).
+- u4: ID-anchor parity with IMP-17 at INSIGHT-MAP §3 line 123. Pure prefix + link; preserves I4 row text.
+
+Out-of-scope (Stage-1 lock GR1-GR6)
+- Any modification to src/renderer.py, src/space_allocator.py, or src/phase_z2_pipeline.py.
+- Role-based ["배경","본심"] hardcoding into Phase Z runtime.
+- Replacing min_height_first + content_weight solver (IMP-09 surface).
+- Migrating ContainerSpec.block_constraints into Phase Z frame contracts.
+- Changing build_containers_type_b call sites.
+- Tests — Stage 1 binding contract precludes source change; docs-only plan with zero runtime surface; pytest baseline already green per Codex Stage 2 R1 (305 passed).
+
+Rollback
+1 unit = 1 commit. u1/u2 → git rm of new doc + revert commit. u3/u4 → git checkout HEAD -- <file>. No src/ touched.
+
+=== EVIDENCE ===
+- src/renderer.py:210-295 consumer pattern verified (Read): OrderedDict grouping by block.area at L218-223; role_order=["배경","본심"] at L234; topic_id + fallback-order match at L240-253; inline-style injection (height_px + font-size + spacing-inner + font-body) at L265-273; len(block_list)==1 single-html vs flex-column branch at L283-291.
+- src/renderer.py:352-353 (render_multi_page with _container_specs) and :426 (render_slide without container_specs) call sites — Stage 1 binding contract.
+- src/space_allocator.py:445-586 build_containers_type_b verified (Read): page_structure → ContainerSpec; zone classification top/bottom/bottom_left/bottom_right/footer at L478-484; slide-base top=65 height=590 at L489-490.
+- src/phase_z2_pipeline.py:794-853 row solver (min_height_first + content_weight_distribution) — Stage 1 binding contract, IMP-09 owned.
+- src/phase_z2_pipeline.py:924-972 col solver (content_weight_distribution_cols) — Stage 1 binding contract.
+- src/phase_z2_pipeline.py:1125-1452 Step 8 dispatch (rows, cols, 2-D dynamic, user override, fr-default fallback) — Stage 1 binding contract.
+- docs/architecture/IMP-18-SVG-GAP-REPORT.md precedent: standalone doc + backlog row marked documented (line 69) — direct template for u1/u3.
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md IMP-19 at line 70, IMP-09 at line 53 (Read).
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md §3 I4 at line 125; IMP-17 ID-prefix precedent at line 123 (Read).
+- pytest baseline already executed by Codex Stage 2 R1: 305 passed in 40.60s. No new pytest required for docs-only plan.
+- Format self-check: no fenced code blocks; no inline code snippets; body order = implementation_units → rationale → out-of-scope → rollback → evidence; total < 5000 chars.

.orchestrator/drafts/19_stage_simulation-plan_codex_r1.md

@@ -0,0 +1,42 @@
+[Codex #3] Stage 2 simulation-plan review R1
+
+Review result: YES.
+
+Plan completeness
+- Issue-body axes covered: renderer._group_blocks_by_area reference extraction, Phase Z Step 8 zone-ratio delta, Phase O no-regression/no-direct-integration guardrails, and soft-link boundary with IMP-09.
+- Stage 1 binding scope respected: no planned edits to src/renderer.py, src/space_allocator.py, or src/phase_z2_pipeline.py; docs/architecture-only reference artifact plus cross-reference updates.
+- Implementation units block exists. u1, u2, u3, and u4 each touch 1 file and estimate <= 50 lines, so the unit-size gate passes.
+- Compact-plan format passes: implementation_units, per-unit rationale, out-of-scope notes, rollback, and evidence are present; no fenced code block or inline code snippet body was found in the plan.
+- Tests are appropriate for a docs-only Stage 3 plan. Baseline pytest was still run for this Stage 2 gate and passed.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- rg -n "def _group_blocks_by_area|OrderedDict|role_order|_group_blocks_by_area|build_containers_type_b|def compute_zone_layout|def compute_zone_layout_cols|IMP-19|IMP-09|I4|IMP-18|IMP-17" src\renderer.py src\space_allocator.py src\phase_z2_pipeline.py docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md
+- Get-ChildItem -LiteralPath docs\architecture -Filter *IMP-18*
+- pytest -q tests
+- Read targeted line windows in src\renderer.py, src\space_allocator.py, src\phase_z2_pipeline.py, docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md, and docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+
+Files checked:
+- src/renderer.py
+- src/space_allocator.py
+- src/phase_z2_pipeline.py
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md
+- docs/architecture/IMP-18-SVG-GAP-REPORT.md
+
+Test results:
+- pytest -q tests: 305 passed in 40.60s
+
+Verified facts:
+- src/renderer.py:210-295 contains the Phase O/Q reference consumer: OrderedDict grouping by block area, role_order for two role containers, topic_id/fallback matching, inline height/font/spacing style injection, and single-block vs flex-column fallback.
+- src/renderer.py:352-353 calls _group_blocks_by_area with page container specs; src/renderer.py:426 calls it without container specs.
+- src/space_allocator.py:445-586 contains build_containers_type_b, including page_structure-driven zone classification and slide-base geometry constants.
+- src/phase_z2_pipeline.py:794-853 contains compute_zone_layout using min_height_first + content_weight_distribution.
+- src/phase_z2_pipeline.py:924-972 contains compute_zone_layout_cols using content_weight_distribution_cols.
+- src/phase_z2_pipeline.py:1125-1452 contains the Step 8 dispatch paths for rows, cols, 2-D dynamic, user override geometry, and fr-default fallback.
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md has the IMP-09 and IMP-19 rows needed for u3 cross-reference edits, and IMP-18 already uses a documented reference-report precedent.
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md has the Step 8 row and the I4 registry row needed for u4 anchor sync.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_simulation-plan_codex_r2.md

@@ -0,0 +1,44 @@
+[Codex #4] Stage 2 simulation-plan review R2
+
+Review result: YES.
+
+Plan completeness
+- Issue-body axes are covered: renderer._group_blocks_by_area reference extraction, Phase Z Step 8 zone-ratio design delta, Phase O no-regression/no-direct-integration guardrails, and the IMP-09 soft-link boundary.
+- The Stage 1 binding scope is respected: planned edits are docs/architecture-only plus cross-reference updates; no edits are planned for src/renderer.py, src/space_allocator.py, or src/phase_z2_pipeline.py.
+- The implementation_units block exists. u1, u2, u3, and u4 each touch 1 file and estimate <= 50 lines, so the unit-size gate passes.
+- The compact-plan format passes: implementation_units, brief per-unit rationale, out-of-scope notes, rollback, evidence, and final marker are present; no fenced code block/code snippet body is used.
+- Tests are appropriate for a docs-only plan, and the required Stage 2 baseline pytest passed.
+
+=== EVIDENCE ===
+Commands run:
+- git status --short
+- rg -n "def _group_blocks_by_area|OrderedDict|role_order|_group_blocks_by_area|build_containers_type_b|def compute_zone_layout|def compute_zone_layout_cols|IMP-19|IMP-09|I4|IMP-18|IMP-17" src\renderer.py src\space_allocator.py src\phase_z2_pipeline.py docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs\architecture\PHASE-Z-PIPELINE-OVERVIEW.md docs\architecture\IMP-18-SVG-GAP-REPORT.md
+- Get-ChildItem -LiteralPath docs\architecture -Filter *IMP-18*
+- Read targeted line windows in src\renderer.py, src\space_allocator.py, src\phase_z2_pipeline.py, docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md, docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md, and docs\architecture\IMP-18-SVG-GAP-REPORT.md
+- pytest -q tests
+
+Files checked:
+- src/renderer.py
+- src/space_allocator.py
+- src/phase_z2_pipeline.py
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md
+- docs/architecture/PHASE-Z-PIPELINE-OVERVIEW.md
+- docs/architecture/IMP-18-SVG-GAP-REPORT.md
+
+Test results:
+- pytest -q tests: 305 passed in 40.61s
+
+Verified facts:
+- src/renderer.py:210-295 contains the Phase O/Q reference consumer: OrderedDict grouping by block area, role_order for two role containers, topic_id/fallback matching, inline height/font/spacing style injection, and single-block vs flex-column fallback.
+- src/renderer.py:352-353 calls _group_blocks_by_area with page container specs; src/renderer.py:426 calls it without container specs.
+- src/space_allocator.py:445-586 contains build_containers_type_b, including page_structure-driven zone classification and slide-base geometry constants.
+- src/phase_z2_pipeline.py:794-853 contains compute_zone_layout using min_height_first + content_weight_distribution.
+- src/phase_z2_pipeline.py:924-972 contains compute_zone_layout_cols using content_weight_distribution_cols.
+- src/phase_z2_pipeline.py:1125-1452 contains the Step 8 dispatch paths for rows, cols, 2-D dynamic, user override geometry, and fr-default fallback.
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md has IMP-09 implemented at line 52, IMP-18 documented at line 68, and IMP-19 pending at line 69, matching the planned cross-reference/status update surface.
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md has IMP-17 with an ID anchor at line 123 and I4 without an IMP-19 anchor at line 125, matching the planned anchor-sync update.
+- docs/architecture/IMP-18-SVG-GAP-REPORT.md is a standalone reference-only docs/architecture precedent with status, scope, source-pattern, activation-gate, and guardrail sections.
+- git status showed only pre-existing untracked workspace files; no tracked source changes were required for this Stage 2 verification.
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_test-verify_claude_r1.md

@@ -0,0 +1,87 @@
+📌 **[Claude #N] Stage 4: 테스트 및 검증 R1 — IMP-19**
+
+■ **검증 결과**
+
+| 축 | 결과 |
+|---|---|
+| pytest -q tests | **305 passed** (Stage 2 R1 baseline 305 와 동일 — regression 0) |
+| `git status --short -- src/` | **clean** — GR1 (no src/ edits) 준수 |
+| `grep -rn "IMP-19" src/` | **0 matches** — GR3 (reference stays in docs/) 준수 |
+| Diff 범위 | 신규 1 (`docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`) + 수정 2 (`PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, `PHASE-Q-INSIGHT-TO-22STEP-MAP.md`) — Stage 2 계획 4 unit 와 정확히 일치 |
+| 하드코딩 | 없음 — role-based `["배경","본심"]` 패턴은 Phase O 원본 (`src/renderer.py:234`) 의 read-only reference 로만 인용; Phase Z 런타임 주입 없음 |
+| Cross-ref 양방향 | ✓ IMP-19 row → reference doc / IMP-09 row → back-ref / INSIGHT-MAP §3 I4 row → IMP-19 anchor + doc link |
+
+■ **Plan ↔ Diff 정합**
+
+- u1 (IMP-19-ZONE-RATIO-REFERENCE.md header + A1 + A2): **OK** — header L1-9, A1 L11-27 (consumer + call sites), A2 L29-44 (producer trace).
+- u2 (A3 + A4 + A5 append): **OK** — A3 L46-70 (Phase Z solver delta + comparison table), A4 L72-81 (IMP-09 boundary), A5 L83-97 (re-activation gate + GR1-GR4).
+- u3 (BACKLOG IMP-09 back-ref + IMP-19 doc link + status pending → documented): **OK** — L53 IMP-09 row appends `soft back-link: IMP-19 ([reference doc](IMP-19-ZONE-RATIO-REFERENCE.md) ...)`, L70 IMP-19 row appends `— [reference doc](...)` + status `documented`.
+- u4 (INSIGHT-MAP §3 I4 row line 125 anchor + doc link): **OK** — `I4 zone 비중 분배` → `IMP-19 I4 zone 비중 분배 (reference — see [IMP-19-ZONE-RATIO-REFERENCE.md](...))`.
+
+총 doc 라인 수: 97 (Stage 2 estimate 74). 초과분은 A3 delta comparison 표 (7-row × 3-col) 가 unit-of-information 단위로 정보 밀도를 높이는 산출물 — Stage 3 Codex verifier 가 이미 accept 함.
+
+■ **File:line 정확도 검증** (random spot-check 6 건)
+
+| Doc claim | Actual code | 일치 |
+|---|---|---|
+| `_group_blocks_by_area` at `src/renderer.py:210-295` | def at L210, return at L295 | ✓ |
+| `role_order = ["배경", "본심"]` at `:234` | L234 exact | ✓ |
+| `font_size` / `padding` defaults at `:262-263` | L262 `block_constraints.get("font_size_px", 15.2)`, L263 `.get("padding_px", 20)` | ✓ |
+| `render_multi_page` call site at `:352-353` | L352-353 with `container_specs=page_container_specs` | ✓ |
+| `render_slide` legacy fallback at `:426` | L426 `_group_blocks_by_area(blocks_raw)` no container_specs | ✓ |
+| `build_containers_type_b` at `src/space_allocator.py:445-586` | def L445, footer-role spec emission ends L588 (logger + return follow at 590-594) — meaningful body captured | ✓ (acceptable; Stage 1 inherited range) |
+
+■ **Guardrails 통과**
+
+- GR1 (no src/ edits) ✓
+- GR2 (Phase O no-regression — `render_slide :426` 변경 없음) ✓
+- GR3 (reference stays in docs/architecture/) ✓
+- GR4 (soft-link integrity — IMP-19 status `documented`, IMP-09 carries back-ref, A5 gate documented) ✓
+- GR5 (IMP-09 boundary preserved — solver ownership not absorbed) ✓
+- GR6 (build_containers_type_b call sites untouched) ✓
+
+■ **결정**: **PASS**. Stage 5 (commit + push) 진입 권고. Commit scope = 3 files (1 new + 2 modified). Suggested message: `docs(IMP-19): zone-ratio reference (Phase O/Q pattern) + IMP-09 back-link`.
+
+=== EVIDENCE ===
+
+```
+commands:
+- git status --short -- docs/architecture/ src/
+- git diff --stat -- docs/architecture/
+- git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md
+- pytest -q tests
+- grep -rn "IMP-19" src/
+- grep -n "IMP-19" docs/architecture/{IMP-19-ZONE-RATIO-REFERENCE.md,PHASE-Q-INSIGHT-TO-22STEP-MAP.md,PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md}
+- wc -l docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md
+
+files_inspected:
+- docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md (97 lines, A1-A5 sections)
+- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md (L53 IMP-09 row, L70 IMP-19 row)
+- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md (L125 I4 row → IMP-19 anchor)
+- src/renderer.py:210-295, 234, 240-253, 261-274, 277-279, 283-291, 352-353, 426
+- src/space_allocator.py:445-594 (function body)
+
+test_results:
+- pytest -q tests = 305 passed in 40.51s (Stage 2 R1 baseline = 305; regression delta = 0)
+
+git_state:
+- src/ clean (no modifications)
+- docs/architecture/ 3 files in working tree (1 untracked + 2 modified, +3/-3 lines combined for modified)
+- HEAD: 02e2ae0
+
+guardrail_verification:
+- GR1 src/-clean: git status --short -- src/ = empty
+- GR2 Phase O render_slide :426 untouched: git diff src/renderer.py = empty
+- GR3 IMP-19 in src/: grep returns 0 matches
+- GR4 status flip pending → documented: confirmed at PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:70
+- GR5 IMP-09 solver untouched: git diff src/phase_z2_pipeline.py = empty
+- GR6 build_containers_type_b call sites: grep -n "build_containers_type_b" src/ untouched
+```
+
+source_comment_ids: [Claude #1] Stage 1 R1; [Claude #3] Stage 2 R2; Stage 3 verifier (Codex Stage 3 R1 accept).
+
+commit_sha: 02e2ae0 (HEAD at Stage 4 entry; commit pending Stage 5).
+
+rewind_target: none (PASS).
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/19_stage_test-verify_codex_r1.md

@@ -0,0 +1,57 @@
+[Codex #N] Stage 4: test-verify R1 — IMP-19
+
+Verification result: PASS. The Stage 3 working tree matches the Stage 2 implementation plan and is safe to commit as a docs-only change.
+
+Scope verification:
+- Changed files are limited to `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`, `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`, and `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`.
+- `src/` is clean in scoped status.
+- `rg -n "IMP-19|IMP-19-ZONE-RATIO-REFERENCE|zone 비중 분배" src/` returned no matches, so the reference axis did not enter runtime code.
+- Existing docs diff is exactly two modified architecture files: backlog + insight-map, with `+3/-3`; the new IMP-19 reference doc is untracked and planned.
+
+Unit verification:
+- u1 OK: `IMP-19-ZONE-RATIO-REFERENCE.md` contains the header plus A1 consumer trace and A2 producer trace.
+- u2 OK: the same doc contains A3 Phase Z solver delta, A4 IMP-09 boundary, and A5 re-activation gate with GR1-GR4.
+- u3 OK: `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` adds an IMP-09 soft back-link to the IMP-19 reference doc and changes IMP-19 status from `pending` to `documented` while adding the reference-doc link.
+- u4 OK: `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` updates the I4 row to carry the IMP-19 anchor and reference-doc link without changing the step or deterministic classification.
+
+Source anchor spot-check:
+- `src/renderer.py:210` defines `_group_blocks_by_area`.
+- `src/renderer.py:234` contains `role_order = ["배경", "본심"]`; this remains read-only reference material and was not introduced into Phase Z.
+- `src/renderer.py:262-263` still applies `font_size_px` and `padding_px` from `block_constraints` in the Phase O/Q container path.
+- `src/renderer.py:353` calls `_group_blocks_by_area(..., container_specs=page_container_specs)` from `render_multi_page`.
+- `src/renderer.py:426` calls `_group_blocks_by_area(blocks_raw)` without `container_specs` from `render_slide`.
+- `src/space_allocator.py:445` defines `build_containers_type_b`.
+- `src/phase_z2_pipeline.py:794` and `src/phase_z2_pipeline.py:924` define the active row and column zone-ratio solvers; dispatch helpers are present at `:1125`, `:1155`, `:1210`, and `:1343`. These remain untouched and IMP-09-owned.
+
+Guardrails:
+- GR1 no `src/` edits: passed.
+- GR2 docs-only reference extraction: passed.
+- GR3 no role-based hardcoding injected into Phase Z runtime: passed.
+- GR4 soft-link integrity: passed; IMP-19 is documented/dormant, with IMP-09 boundary text and backlink.
+- GR5 IMP-09 solver ownership preserved: passed.
+- GR6 `build_containers_type_b` call sites untouched: passed.
+
+=== EVIDENCE ===
+Commands run:
+- `git status --short -- docs/architecture/ src/`
+- `git diff --stat -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md -- docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `rg -n "IMP-19|zone ratio|_group_blocks_by_area|build_containers_type_b|compute_zone_layout|compute_zone_layout_cols|role_order|reference doc|documented|I4" docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md src/renderer.py src/space_allocator.py src/phase_z2_pipeline.py`
+- `pytest -q tests`
+- `git status --short -- src/`
+- `rg -n "IMP-19|IMP-19-ZONE-RATIO-REFERENCE|zone 비중 분배" src/`
+- `Select-String -LiteralPath 'src/renderer.py' -Pattern '^def _group_blocks_by_area','role_order =','font_size =','padding =','container_specs=page_container_specs','_group_blocks_by_area\(blocks_raw\)'`
+- `Select-String -LiteralPath 'src/space_allocator.py' -Pattern '^def build_containers_type_b','return containers'`
+- `Select-String -LiteralPath 'src/phase_z2_pipeline.py' -Pattern '^def compute_zone_layout\(','^def compute_zone_layout_cols\(','^def _build_rows_dynamic','^def _build_grid_dynamic_2d','^def _build_cols_dynamic','^def build_layout_css'`
+- `(Get-Content -LiteralPath 'docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md').Count; git ls-files --others --exclude-standard -- docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+Files checked:
+- `docs/architecture/IMP-19-ZONE-RATIO-REFERENCE.md`
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- `src/renderer.py`
+- `src/space_allocator.py`
+- `src/phase_z2_pipeline.py`
+Test results:
+- `pytest -q tests` => 305 passed in 40.44s
+
+FINAL_CONSENSUS: YES

.orchestrator/drafts/20_stage_code-edit_claude_r1.md

@@ -0,0 +1,55 @@
+Executing unit: u1
+
+📌 **[Claude #1] Stage 3 Round #1 — unit u1 executed**
+
+■ **unit_executed**
+- `u1` — Create `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` with header + § A1 Phase Q consumer reference + § A2 `REQUIRED_PATTERNS` shape table.
+
+■ **files_changed**
+- `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` (new, 48 lines)
+
+■ **diff_summary**
+- New file created. Untracked per `git status --short`: `?? docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`.
+- Header block (L1-9): status=`documented (reference-only, dormant)`, scope=`doc-only`, related issue link, soft dependency `IMP-04` with A5-gate caveat, source axis `INSIGHT-MAP §3 / §2.7 H2`. Mirrors IMP-19 reference doc header structure (precedent).
+- § A1 Phase Q consumer pattern (read-only reference), L11-22:
+  - `src/content_verifier.py:382-392` — `REQUIRED_PATTERNS` dict + verified values (`body_bg=[]`, `body_core=["key-msg"]`, `sidebar=["padding-left","text-indent"]`, `footer=[]`) + Phase T `L379-381` comment no-regression boundary (GR2 from Stage 2).
+  - `src/content_verifier.py:395-448` — `verify_structure` substring + OR (`pattern.split("|")` at L410) + `has_image` branch (L414-416, `"slide-img-"` injection only when `area=="body_core"`) + `font_hierarchy` 1px tolerance (L433 `fs > max_font + 1`) + result construction (L438 binary `passed`, L445 continuous score `1.0 - len(missing)/max(1, len(patterns))`).
+  - `src/content_verifier.py:455-487` — `verify_area` L1+L2+L3 composition; L465 `verify_structure` call passes `has_image` but **not** `font_hierarchy`.
+  - `src/content_verifier.py:490-529` — `verify_all_areas` dispatch; L518 confirmed body_core is the only branch propagating `has_image`.
+  - Classification line: area-level (Phase Q HTML area axis), **not** frame_id × sub_zone.
+- § A2 `REQUIRED_PATTERNS` shape table (read-only reference), L26-46:
+  - 7-row axis table: key axis (area string) / value type (`list[str]`) / alternation (`|` OR) / image-conditional branch / font-tolerance (1px) / pass-score rule / empty-pattern handling.
+  - Shape-only carry-over candidates listed (5 items: dict indirection, OR semantics, conditional implicit requirement, continuous score degradation, separate errors vs warnings lanes).
+  - Explicit exclusion list at bottom: literal strings `"key-msg"`, `"padding-left"`, `"text-indent"`, `"slide-img-"` and area names `body_bg`/`body_core`/`sidebar`/`footer` **must not** carry into Phase Z (Phase Q area-HTML idioms, not Phase Z frame/slot idioms) — enforces GR1 (shape-only, no value reuse) from Stage 2.
+
+■ **scope_lock_adherence**
+- GR1 (shape-only, no value reuse) — § A2 final paragraph explicitly excludes literal pattern strings and area names from carry-over.
+- GR2 (Phase Q REQUIRED_PATTERNS no-touch) — § A1 acknowledges Phase T `L379-381` comment as preserved no-regression boundary; no edits to `src/content_verifier.py`.
+- GR3 (Phase Z dict = Phase Z-owned) — deferred to u2 (§ A3 target design).
+- GR4 (IMP-04 soft-link one-way) — header line cites soft dep but defers boundary statement to u2 (§ A4).
+- PZ-1 (no AI/Kei rewrite) — pure spec doc, no AI invocation.
+- RULE 12 (full paths) — all file refs use absolute repo-rooted paths (`src/content_verifier.py:NNN`).
+- 1 turn = 1 unit — u2/u3/u4 deferred.
+
+■ **remaining_units**
+- `u2` — Append § A3 (Phase Z frame_id × sub_zone target dict + abstraction-mismatch table) + § A4 (IMP-04 soft-link boundary) + § A5 (re-activation gate 3-cond AND + GR1–GR4 + Q1–Q4 resolutions) to the same doc (~48 lines).
+- `u3` — Update `PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` IMP-20 row L71 (`pending` → `documented` + ref-doc link) and `PHASE-Z-PIPELINE-STATUS-BOARD.md` Step 10 row L49 (append IMP-20 ref-doc link; ⚠ partial unchanged).
+- `u4` — Update `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 H2 row L126 (prepend `IMP-20 ` + append `(reference — see [IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md])`).
+
+■ **follow_up_issue_candidates**
+- None this turn. Scope-lock 외 axis not encountered during u1 execution. All references remained within `src/content_verifier.py:382-529` as scoped in Stage 2 `evidence` block.
+
+■ **next**
+- Codex review of u1 → continue_same for u2 (per Stage 3 unit progression convention).
+- No commit, no push (Stage 3 rule — Gitea comment is the deliverable).
+
+```yaml
+unit_executed: u1
+files_changed:
+  - docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md
+remaining_units: [u2, u3, u4]
+follow_up_issue_candidates: []
+stage: 3_code_edit
+turn: 1
+commit_sha: null  # no commit per Stage 3 rule
+```

.orchestrator/drafts/20_stage_code-edit_claude_r2.md

@@ -0,0 +1,47 @@
+Executing unit: u2
+
+📌 **[Claude #2] Stage 3 Round #2 — unit u2 executed**
+
+■ **unit_executed**
+- `u2` — Append `§ A3` (Phase Z target pattern dict against `frame_contracts.yaml` 11 contracts + abstraction-mismatch table) + `§ A4` (IMP-04 soft-link boundary, catalog vs validation ownership) + `§ A5` (re-activation gate 3-cond AND + Q1–Q4 resolved + GR1–GR4 + PZ-1 + RULE 13) to `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`.
+
+■ **files_changed**
+- `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` (existing untracked, 48 → 109 lines, +61 lines appended)
+
+■ **diff_summary**
+- Doc-only append. `git status --short` confirms the file is still the single untracked artifact; no tracked-source diff (no `src/`, no `templates/phase_z2/catalog/frame_contracts.yaml`, no backlog / status-board / INSIGHT-MAP edits — those land in u3/u4).
+- § A3 (Phase Z target pattern dict, design input only), enumerated all 11 frame contracts with file:line anchors:
+  - `templates/phase_z2/catalog/frame_contracts.yaml:21` F13 `three_parallel_requirements` (3 sub_zones), `:77` F29 `process_product_two_way` (2 × strict 3), `:128` F16 `bim_issues_quadrant_four` (4), `:189` F14 `three_persona_benefits` (3), `:253` F12 `construction_goals_three_circle_intersection` (3+1, intersection `min:0,max:1`), `:323` F11 `construction_bim_three_usage` (3), `:391` F18 `bim_dx_comparison_table` (2 header + rows `min:1,max:12`), `:456` F20 `dx_sw_necessity_three_perspectives` (3), `:520` F8 `info_management_what_how_when` (3), `:580` F28 `sw_reality_three_emphasis` (3), `:637` F17 `bim_current_problems_paired` (8, row×side).
+  - 11/11 carry `accepted_content_types` + `sub_zones`; `density_envelope` absent (`grep -c "density_envelope" templates/phase_z2/catalog/frame_contracts.yaml` = 0).
+  - Active-surface anchors: `src/phase_z2_mapper.py:49-57` `load_frame_contracts` / `get_contract` (dict lookup); `src/phase_z2_pipeline.py:3776-3805` Step 10 partial emit (`step_status="partial"`, no pattern-dict assertion yet).
+  - Abstraction-mismatch table — 7 axis rows (Key / Cardinality of keys / Value semantics / Conditional branch input / Tolerance / Validation timing / Result lanes) Phase Q (area axis) vs Phase Z target (`(frame_id, sub_zone_id)` tuple axis). Key cardinality: 4 fixed area names (Phase Q) vs 39 sub_zones across current catalog (3+2+4+3+4+3+3+3+3+3+8) (Phase Z). Final paragraph reinforces "shape-only portable; every value is Phase Q-local" — locks GR1.
+- § A4 (IMP-04 soft-link boundary, catalog vs validation ownership):
+  - Backlog anchor `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md:71` cited as soft-link source.
+  - Ownership split: **IMP-04 owns** every `frame_contracts.yaml` entry / `accepted_content_types` / `sub_zones` / `cardinality` / `visual_hints` mutation. **IMP-20 owns** only A1+A2+A3 reference narrative.
+  - Coupling direction = **one-way read** (Phase Z dict consumes catalog; never publishes back; IMP-04 unaware of IMP-20). Reference direction = one-way (no reverse pointer in source files). Schema-change clause: IMP-04 catalog schema change ⇒ A3 re-verify, A4 statement unchanged.
+- § A5 (re-activation gate + design Q resolution + guardrails):
+  - **3-cond AND**: (1) Step 10 partial-emit-insufficient regression that maps to frame/slot axis (not composition / adapter / render-CSS layer); (2) failing-case MDX + `step10_frame_contract.json` trace + final.html slot-path excerpt; (3) IMP-04 owner sign-off that catalog tightening alone cannot resolve.
+  - **Q1 → `(frame_id, sub_zone_id)`** (frame-only insufficient — F29 has heterogeneous `accepts`: `process_column=[text_block, transform_table]` vs `product_column=[text_block]`).
+  - **Q2 → hybrid (substring + contract-field + tolerance, 3 independent lanes)**.
+  - **Q3 → post Step 18 final.html only** (Step 12 light_edit/restructure exposes no DOM; running dict there → false negatives or duplicate work).
+  - **Q4 → font-hierarchy carry-over: shape portable (1px tolerance), source replaced** (Phase Q `role_font_map` fixed → Phase Z `visual_hints` catalog-driven; future `max_font_px` field IMP-04-owned).
+  - **Guardrails**: GR1 (shape-only — no `"key-msg"` / `"padding-left"` / `"text-indent"` / `"slide-img-"` / area name carry); GR2 (Phase Q `content_verifier.py:382-392` no-touch, Phase T `L379-381` overflow:hidden boundary preserved); GR3 (Phase Z dict Phase Z-owned — no `import REQUIRED_PATTERNS`); GR4 (IMP-04 soft-link one-way); PZ-1 (AI isolation — dict is code/spec, no Kei rewrite); RULE 13 (backlog ↔ status-board ↔ INSIGHT-MAP same-commit sync on any future activation).
+
+■ **scope_lock_adherence**
+- GR1 (shape-only, no value reuse) — A3 abstraction-mismatch table contrasts axes only; A5 GR1 restated literal-string exclusion list; no Phase Q literal carries into the Phase Z design target.
+- GR2 (Phase Q `REQUIRED_PATTERNS` no-touch) — `src/content_verifier.py` untouched (`git status --short` confirms no tracked diff). A1 (u1) + A5 GR2 both restate Phase T `L379-381` no-regression boundary.
+- GR3 (Phase Z dict = Phase Z-owned) — A5 GR3 explicit: no `import` of `content_verifier.REQUIRED_PATTERNS` from Phase Z; symbol isolation preserved.
+- GR4 (IMP-04 soft-link one-way) — § A4 codifies one-way read direction; § A5 GR4 restates non-blocking semantics (IMP-20 activation must not block on / modify IMP-04).
+- PZ-1 (AI isolation contract) — § A5 PZ-1 line: pattern dict is code/spec, not AI-generated. Consistent with `feedback_ai_isolation_contract`.
+- RULE 13 (anchor sync) — A5 commitment recorded; actual backlog (L71 → `documented`) + status-board (L49 IMP-20 ref-doc link append) updates land in u3, INSIGHT-MAP §3 H2 row update in u4 (per Stage 2 plan).
+- Out-of-scope preserved: no Step 10 partial → complete promotion, no IMP-04 implementation, no `frame_contracts.yaml` schema change, no `src/content_verifier.py` edit, no AI/Kei call, no tests, no commit.
+
+■ **remaining_units**
+- `[u3, u4]`
+  - **u3**: PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md L71 (`pending` → `documented` + ref-doc link) + PHASE-Z-PIPELINE-STATUS-BOARD.md Step 10 row L49 append IMP-20 ref-doc link.
+  - **u4**: PHASE-Q-INSIGHT-TO-22STEP-MAP.md §3 H2 row L126 — prepend `IMP-20 ` + append `(reference — see [IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md])`. Mirrors IMP-19 L125 / IMP-17 L123 precedent.
+
+■ **follow_up_issue_candidates**
+- None new from this unit. The four Stage 2 follow_up_candidates carry forward unchanged: (a) § A5 3-cond AND gate evaluation (separate axis); (b) Step 10 partial → complete promotion (gated on § A5); (c) IMP-04 implementation surface (gated on § A4 soft-link confirmation); (d) Phase Z runtime pattern-dict activation (gated on § A5). All four remain outside this Stage 3 scope.
+
+STOP. No commit, no push. Awaiting Codex r2 verification of u2 (Gitea comment is the deliverable).

.orchestrator/drafts/20_stage_code-edit_claude_r3.md

@@ -0,0 +1,46 @@
+Executing unit: u3
+
+📌 **[Claude #3] Stage 3 Round #3 — unit u3 executed**
+
+■ **unit_executed**
+- `u3` — Bidirectional cross-ref sync: (a) backlog `IMP-20` row (`pending` → `documented` + scope-cell ref-doc link) and (b) status-board Step 10 row (append IMP-20 ref-doc link; ⚠ partial unchanged).
+
+■ **files_changed**
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (1-line edit at L71)
+- `docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` (1-line edit at L49)
+
+■ **diff_summary**
+- Backlog L71 (`docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`):
+  - scope cell: appended ` — [reference doc](IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md)` after the existing scope text (`content_verifier.verify_structure pattern reference — Phase Z frame contract 검증 pattern`). Mirrors IMP-19 row L70 link convention (`— [reference doc](IMP-19-ZONE-RATIO-REFERENCE.md)`) and IMP-18 row L69 link convention (`— [gap report](IMP-18-SVG-GAP-REPORT.md)`).
+  - status cell: `pending` → `documented`. Matches IMP-18 / IMP-19 sibling rows.
+  - Other cells unchanged: title (`H2 frame contract validation`), related layer (`Step 10`), source (`§3 Reference Only`), priority (`↓ low`), guardrail (`Phase Q REQUIRED_PATTERNS 값 회귀 X / Phase Z 자체 pattern dict 설계`), dependency (`soft link: IMP-04`). GR1 / GR2 / GR3 / GR4 are preserved in-cell as the existing guardrail / dependency text — no edits.
+- Status-board L49 (`docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md`):
+  - Step 10 row Status cell: appended ` — IMP-20 ref: [reference doc](IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md)` after the existing `(B3 의 accepted_content_types + sub_zones 선언 추가 — B4 만 읽음, mapper 미읽음 / density envelope 별 axis)` partial annotation.
+  - `⚠ partial` symbol and existing partial-status text remain unchanged — Stage 2 plan u3 lock (`⚠ partial unchanged`) honored. Codex r1 line-number fix folded in (L49, not L86).
+- `git diff --` on both files shows the 2 single-line replacements (one per file) and nothing else; CRLF warnings are git's normal Windows newline notice. No `src/`, no `templates/`, no `INSIGHT-MAP` edits in this unit (u4 owns INSIGHT-MAP).
+- Total tracked-source diff in u3 = 2 lines (-) / 2 lines (+). No new files. No deletions.
+
+■ **scope_lock_adherence**
+- GR1 (shape-only, no value reuse): u3 makes no changes to Phase Q literal patterns or Phase Z dict design — pure cross-ref sync. No regression.
+- GR2 (Phase Q `REQUIRED_PATTERNS` no-touch): `src/content_verifier.py` unchanged in u3 (confirmed by scoped `git diff --` on the file path — empty).
+- GR3 (Phase Z dict = Phase Z-owned): u3 only references the existing `IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` (u1+u2 doc) — no new Phase Q import path introduced anywhere.
+- GR4 (IMP-04 soft-link one-way only): backlog dependency cell already encodes `soft link: IMP-04 (확장 catalog 적용 시 검증 범위 확대)` — preserved verbatim in u3, no coupling change.
+- PZ-1 (no AI/Kei rewrite): doc-only sync — zero AI surface.
+- RULE 13 (anchor sync — bidirectional): backlog `documented` status now matches sibling IMP-18 / IMP-19 sibling rows; status-board Step 10 row now contains a back-pointer to the reference doc. The INSIGHT-MAP §3 H2 row (third anchor leg) lands in u4 — explicitly out-of-scope for u3 per Stage 2 plan.
+- PZ-2 (1 turn = 1 step): only u3 executed in this turn. u4 not started.
+
+■ **scope_qualified verification**
+- Scope: Stage 3 u3 only — backlog L71 row + status-board L49 row.
+- Backlog L71 row verified after edit: status cell reads `documented`; scope cell ends with `— [reference doc](IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md)`.
+- Status-board L49 row verified after edit: row still begins with `| A | 10 | Frame Contract 확인 | ⚠ partial` and ends with `— IMP-20 ref: [reference doc](IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md) |`.
+- Out-of-scope (not verified in this unit): INSIGHT-MAP §3 H2 row at L126 (owned by u4); `frame_contracts.yaml`; `src/content_verifier.py`; `src/phase_z2_pipeline.py`; `src/phase_z2_mapper.py`.
+
+■ **remaining_units**
+- `u4` — Update `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 H2 row at L126: prepend `IMP-20 ` + append `(reference — see [IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md])`. Mirrors IMP-19 L125 / IMP-17 L123 precedent.
+
+■ **follow_up_issue_candidates**
+- None new in this unit. The 4 follow-up axes already documented in u2 § A5 (3-cond AND gate evaluation, Step 10 partial → complete promotion, IMP-04 implementation activation, Phase Z runtime pattern-dict activation) remain dormant per Stage 2 plan and require separate axis triggers — explicitly NOT actioned here.
+
+■ **next**
+- STOP. No commit, no push. Per Stage 3 contract, the Gitea comment is the deliverable.
+- Awaiting Codex r3 review of u3, then proceeding to u4 in the next turn (1 turn = 1 unit).

.orchestrator/drafts/20_stage_code-edit_claude_r4.md

@@ -0,0 +1,55 @@
+Executing unit: u4
+
+📌 **[Claude #4] Stage 3 Round #4 — unit u4 executed**
+
+■ **unit_executed**
+- `u4` — `PHASE-Q-INSIGHT-TO-22STEP-MAP.md` §3 H2 row L126 sync: prepend `IMP-20 ` to the axis cell + append `(reference — see [`IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`](IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md))` to the same axis cell. Mirrors L125 IMP-19 (`reference` keyword variant) / L123 IMP-17 (`carve-out` keyword variant) precedent.
+
+■ **files_changed**
+- `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (1-line edit at L126)
+
+■ **diff_summary**
+- L126 axis cell BEFORE: `H2 frame contract validation`
+- L126 axis cell AFTER: `IMP-20 H2 frame contract validation (reference — see [`IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`](IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md))`
+- Other cells UNCHANGED:
+  - 관련 step: `Step 10`
+  - source: ``§2.7 H2 (`content_verifier.verify_structure` pattern)``
+  - status: `pending` (table-local status; precedent — L123 IMP-17 / L125 IMP-19 both keep `pending` in INSIGHT-MAP table even after their reference docs were issued, because INSIGHT-MAP is a *mapping* doc; backlog L71 + status-board L49 carry the canonical `documented` / `partial` status — u3 already synced those).
+  - normal path 여부: `yes (deterministic)`
+- `git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` = 1 line removed (-) / 1 line added (+); CRLF warning is git's normal Windows newline notice. No other content shifted.
+- Total tracked-source diff in u4 = 1 line (-) / 1 line (+). No new files, no deletions.
+
+■ **scope_lock_adherence**
+- GR1 (shape-only, no value reuse): u4 is INSIGHT-MAP row sync only — no Phase Q literal pattern or Phase Z dict design surface touched.
+- GR2 (Phase Q `REQUIRED_PATTERNS` no-touch): `src/content_verifier.py` UNCHANGED in u4 (confirmed by scoped `git diff --` on the runtime path — empty).
+- GR3 (Phase Z dict = Phase Z-owned): u4 only forward-links to the existing `IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` doc (u1 + u2 product) — no new Phase Q import path created.
+- GR4 (IMP-04 soft-link one-way only): u4 makes no IMP-04 reference; the IMP-20 ref-doc itself (`§ A4` from u2) carries the soft-link clause one-way. No coupling change.
+- PZ-1 (no AI/Kei rewrite): doc-only sync — zero AI surface.
+- RULE 13 (bidirectional cross-ref sync): u4 closes the third leg of the 3-leg sync — backlog (u3) ↔ status-board (u3) ↔ INSIGHT-MAP (u4). All three docs now cross-reference `IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` in their canonical IMP-20 row / Step 10 row / §3 H2 row.
+- RULE 12 (full paths): all links use repo-relative paths consistent with L123 / L125 precedent (no absolute `D:\...` because these are repository markdown anchors, not user-facing file references).
+
+■ **remaining_units**
+- `[]` (all 4 implementation_units from Stage 2 plan now executed: u1 + u2 + u3 + u4).
+
+■ **follow_up_issue_candidates**
+- None within this stage's scope. Out-of-scope follow-up axes already enumerated in Stage 2 `follow_up_candidates`:
+  - § A5 3-cond AND gate evaluation (gated separate axis).
+  - Step 10 partial → complete promotion (gated on § A5).
+  - IMP-04 implementation (gated on § A4 soft-link confirmation).
+  - Phase Z runtime pattern-dict activation (gated on § A5).
+  - These are documented in `IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` § A5 (re-activation gate) — no new follow-up issue needed in this Stage 3 round.
+
+■ **evidence**
+- `git diff -- docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` → shows L126 single-line replacement (axis cell prepend + append), other cells (`Step 10` / source / `pending` / `yes (deterministic)`) UNCHANGED.
+- `git diff -- src/content_verifier.py src/phase_z2_pipeline.py src/phase_z2_mapper.py templates/phase_z2/catalog/frame_contracts.yaml` → empty (no runtime / catalog surface touched in u4).
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` → shows only the u3 diffs (already verified in r3) — no u4 leakage into u3-owned files.
+- `git status --short` (tracked subset) → shows exactly 3 modified files for the full unit chain: `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` (M, u4), `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` (M, u3), `docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` (M, u3). Untracked artifact: `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` (??, u1 + u2). Untracked test-matching / orchestrator / sample artifacts are unrelated to IMP-20.
+- Precedent verification — L123 IMP-17 / L125 IMP-19 both show `IMP-NN <original-axis-name> (<keyword> — see [<ref-doc>.md](<ref-doc>.md))` axis-cell pattern + table-local `pending` status retained. u4 follows L125 IMP-19 `reference` keyword variant (matches u1 header status `documented (reference-only, dormant)`).
+
+■ **commit_sha**
+- N/A — Stage 3 = no commit, no push per orchestrator stage charter. Stage 5 (commit-push) is a separate gate.
+
+■ **end-of-Stage-3 readiness**
+- All 4 implementation_units complete (u1 ✓ Claude r1 + Codex r1 verified → u2 ✓ Claude r2 + Codex r2 verified → u3 ✓ Claude r3 + Codex r3 verified → u4 ✓ executed this round, awaiting Codex r4 verification).
+- After Codex r4 `FINAL_CONSENSUS: YES`, the next stage is Stage 4 (test-verify) — but `test_plan: None` per Stage 2 plan (docs-only, no runtime / no `pytest`). Stage 4 may collapse into a docs-only spot-check or direct progression to Stage 5 (commit-push) per orchestrator stage policy; Claude r4 does not advance scope here, per `feedback_one_step_per_turn`.
+- STOP.

.orchestrator/drafts/20_stage_code-edit_codex_r1.md

@@ -0,0 +1,35 @@
+[Codex #1] Stage 3 Round #1 - unit u1 verification
+
+unit_verified: u1
+
+Verdict: correct for the single executed unit. The new `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` file exists, is doc-only, and contains the required header plus A1/A2 Phase Q reference extraction. I verified the cited source anchors against `src/content_verifier.py`; no Phase Q runtime logic, Phase Z runtime logic, `frame_contracts.yaml`, backlog, status board, or INSIGHT-MAP file was modified in this unit.
+
+Verified scope:
+- Header block present with status `documented (reference-only, dormant)`, doc-only scope, issue link, IMP-04 soft dependency, and INSIGHT-MAP H2 source axis.
+- A1 documents the Phase Q consumer reference only: `REQUIRED_PATTERNS`, `verify_structure`, `verify_area`, and `verify_all_areas`.
+- A2 documents the reusable shape only: dict indirection, list values, `|` OR semantics, conditional `has_image`, 1px font tolerance, pass/score behavior, and empty-pattern handling.
+- The document explicitly excludes carrying Phase Q literal strings and area names into Phase Z, satisfying the shape-only guardrail.
+- `git status --short` shows the new doc as untracked and no tracked source diff for `src/content_verifier.py`.
+
+=== EVIDENCE ===
+Commands run:
+- `Get-Content -LiteralPath 'D:\ad-hoc\kei\design_agent\docs\architecture\IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md'`
+- `rg -n "REQUIRED_PATTERNS|def verify_structure|pattern.split|slide-img-|font_hierarchy|fs > max_font|def verify_area|verify_structure\(|def verify_all_areas|body_core|overflow:hidden|overflow" src/content_verifier.py`
+- `git status --short`
+- `$i=1; Get-Content -LiteralPath 'docs\architecture\IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md' | ForEach-Object { '{0,4}: {1}' -f $i, $_; $i++ }`
+- `$i=1; Get-Content -LiteralPath 'src\content_verifier.py' | ForEach-Object { if ($i -ge 379 -and $i -le 531) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `git diff -- docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md src/content_verifier.py`
+- `rg -n -C 8 "^REQUIRED_PATTERNS|^def verify_structure|^def verify_area|^def verify_all_areas" src/content_verifier.py`
+- `rg -n -C 4 'pattern.split|slide-img-|font_warnings|score=1\.0|verify_structure\(|body_html|has_image="body_core"|sidebar_html|footer_html' src/content_verifier.py`
+- `rg -n "^# IMP-20|^## A1|^## A2|Values that|Shape-only|dict\[key\]|Phase Q implements|src/content_verifier.py:382|src/content_verifier.py:395|src/content_verifier.py:455|src/content_verifier.py:490" docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+
+Test results: not run; current stage is single-unit docs verification and explicitly does not require full pytest.
+
+Remaining units: [u2, u3, u4]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/20_stage_code-edit_codex_r2.md

@@ -0,0 +1,45 @@
+[Codex #2] Stage 3 Round #2 - unit u2 verification
+
+unit_verified: u2
+
+Verdict: correct for the single executed unit. The existing `docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` document now contains the required A3, A4, and A5 sections, and the change remains doc-only for this unit. I verified the cited Phase Z catalog, mapper, and Step 10 pipeline anchors. No Phase Q runtime logic, Phase Z runtime logic, `frame_contracts.yaml`, backlog, status board, or INSIGHT-MAP file was modified by this unit.
+
+Verified scope:
+- A3 defines the Phase Z target as a `frame_id x sub_zone` pattern-dict axis aligned to `templates/phase_z2/catalog/frame_contracts.yaml`.
+- A3 lists all 11 active catalog contracts and the sub-zone counts match the parsed catalog: 3, 2, 4, 3, 4, 3, 3, 3, 3, 3, and 8, for 39 total sub-zones.
+- All 11 contracts have `accepted_content_types` and `sub_zones`; `density_envelope` is absent from the catalog.
+- `src/phase_z2_mapper.py:49-57` is correctly cited as the `load_frame_contracts` / `get_contract` direct lookup surface.
+- `src/phase_z2_pipeline.py:3776-3805` is correctly cited as the Step 10 partial emit surface, including `accepted_content_types`, `sub_zones`, payload-builder fields, and `step_status="partial"`.
+- A4 records the IMP-04 soft-link boundary as catalog ownership vs IMP-20 validation-reference ownership, with one-way read coupling.
+- A5 records the dormant re-activation gate, resolves Q1-Q4, and preserves GR1-GR4 plus PZ-1 and RULE 13.
+- Correction to Claude r2 wording: `git status --short` does not show this document as the single untracked artifact in the whole worktree; there are many unrelated untracked files. The relevant scoped check is that `git diff --` for the runtime/catalog/status-board/INSIGHT-MAP paths is empty.
+
+=== EVIDENCE ===
+Commands run:
+- `$i=1; Get-Content -LiteralPath 'docs\architecture\IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md' | ForEach-Object { if ($i -ge 1 -and $i -le 140) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `git status --short`
+- `rg -n "^F[0-9]+:|accepted_content_types|sub_zones|density_envelope" templates/phase_z2/catalog/frame_contracts.yaml`
+- `$i=1; Get-Content -LiteralPath 'src\phase_z2_mapper.py' | ForEach-Object { if ($i -ge 45 -and $i -le 60) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `$i=1; Get-Content -LiteralPath 'src\phase_z2_pipeline.py' | ForEach-Object { if ($i -ge 3770 -and $i -le 3810) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `rg -n "three_parallel_requirements|process_product_two_way|bim_issues_quadrant_four|three_persona_benefits|construction_goals_three_circle_intersection|construction_bim_three_usage|bim_dx_comparison_table|dx_sw_necessity_three_perspectives|info_management_what_how_when|sw_reality_three_emphasis|bim_current_problems_paired" templates/phase_z2/catalog/frame_contracts.yaml`
+- `Select-String -LiteralPath 'templates\phase_z2\catalog\frame_contracts.yaml' -Pattern 'density_envelope' | Measure-Object | ForEach-Object { $_.Count }`
+- `$i=1; Get-Content -LiteralPath 'docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md' | ForEach-Object { if ($i -ge 66 -and $i -le 74) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `git diff -- docs/architecture/IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md src/content_verifier.py src/phase_z2_pipeline.py src/phase_z2_mapper.py templates/phase_z2/catalog/frame_contracts.yaml docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+- Python YAML parse of `templates/phase_z2/catalog/frame_contracts.yaml` to count contracts and sub-zones.
+- `rg -n "visual_hints|min_height_px|max_font_px|strict:|accepts:|partial_target_path:" templates/phase_z2/catalog/frame_contracts.yaml`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_mapper.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-STATUS-BOARD.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+
+Test results: not run; current stage is single-unit docs verification and explicitly says not to run full pytest.
+
+Remaining units: [u3, u4]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO

.orchestrator/drafts/20_stage_code-edit_codex_r3.md

@@ -0,0 +1,34 @@
+[Codex #3] Stage 3 Round #3 - unit u3 verification
+
+unit_verified: u3
+
+Verdict: correct for the single executed unit. The u3 change is limited to the planned cross-reference sync in the backlog and status board. The backlog IMP-20 row now links to `IMP-20-FRAME-CONTRACT-VALIDATION-REFERENCE.md` and has status `documented`; the Step 10 status-board row keeps `partial` status and appends the IMP-20 reference-doc link.
+
+Verified scope:
+- `docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md` changes only the IMP-20 row: scope cell gains the reference-doc link and status changes from `pending` to `documented`.
+- `docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md` changes only the Step 10 row: the existing partial-status text is preserved and the IMP-20 reference-doc link is appended.
+- The guardrail and dependency cells on the IMP-20 backlog row remain unchanged, preserving the Phase Q no-touch guardrail, the Phase Z-owned pattern-dict note, and the IMP-04 soft-link boundary.
+- The scoped diff for runtime/catalog/INSIGHT-MAP files is empty: `src/content_verifier.py`, `src/phase_z2_pipeline.py`, `src/phase_z2_mapper.py`, `templates/phase_z2/catalog/frame_contracts.yaml`, and `docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md` were not modified by this unit.
+
+=== EVIDENCE ===
+Commands run:
+- `$i=1; Get-Content -LiteralPath 'docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md' | ForEach-Object { if ($i -ge 66 -and $i -le 74) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `$i=1; Get-Content -LiteralPath 'docs\architecture\PHASE-Z-PIPELINE-STATUS-BOARD.md' | ForEach-Object { if ($i -ge 45 -and $i -le 52) { '{0,4}: {1}' -f $i, $_ }; $i++ }`
+- `git diff -- docs/architecture/PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md docs/architecture/PHASE-Z-PIPELINE-STATUS-BOARD.md`
+- `git diff -- src/content_verifier.py src/phase_z2_pipeline.py src/phase_z2_mapper.py templates/phase_z2/catalog/frame_contracts.yaml docs/architecture/PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+
+Files checked:
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-IMPLEMENTATION-ISSUE-BACKLOG.md`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Z-PIPELINE-STATUS-BOARD.md`
+- `D:\ad-hoc\kei\design_agent\src\content_verifier.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_pipeline.py`
+- `D:\ad-hoc\kei\design_agent\src\phase_z2_mapper.py`
+- `D:\ad-hoc\kei\design_agent\templates\phase_z2\catalog\frame_contracts.yaml`
+- `D:\ad-hoc\kei\design_agent\docs\architecture\PHASE-Q-INSIGHT-TO-22STEP-MAP.md`
+
+Test results: not run; current stage is single-unit docs verification and explicitly does not require full pytest.
+
+Remaining units: [u4]
+
+rewind_target: continue_same
+FINAL_CONSENSUS: NO