1. V4 evidence 확장 (pipeline_17b_extend_missing_sections.py):
'결과물이 아니라 프로세스' 원칙 — pipeline_17 과 동일 평가 코드로 누락 3개
섹션(01-intro/05-1/05-2)만 평가해 v4_full32_result.yaml 병합 (기존 무접촉,
blind/ANSWER_MAP 불변). 결과: 05-1 → F20 light_edit 0.77 (design-matched!),
01-intro/05-2 → all-reject (catalog gap 정직 노출 — F19 가 05 주제와 이름까지
일치하나 partial 없음 → #2 프로모션 최우선 근거)
2. renderable-aware provisional (IMP-30 u1 정밀화):
rank-1 무조건 승격 → partial 존재 AND (비-reject OR verbatim builder 보유)
첫 후보 승격. reject+builder 미보유 renderable 의 mapper 네이티브 렌더는
원문 drop 위험 (F23 1-atom 손실 실측) — 원문 보존 > design 개선 우선순위.
3. 게이트 순서 버그 수정 (_apply_quality_gate_downgrades 추출):
T28.5d popup 승격 후 재계산 경로에 quality gate 강등 3종(coverage/forbidden/
consistency) 미적용 → 텍스트 손실이 overall=PASS 로 통과 (mdx05 실측).
양 경로 공통 헬퍼로 통일 — mdx04 의 #16 시점 PASS 일부가 이 버그 덕이었음을
정직하게 정정 (현재 PARTIAL + frame mismatch 라벨, 텍스트는 완전).
최종 5-MDX: 전부 missing_atoms=0 (무손실) / 01·03 PASS / 02·04·05 PARTIAL(정직
frame-mismatch 라벨) / mdx01 readiness not_ready→needs_review 개선.
게이트: 1061 passed. SHA baseline 재캡처 (정당 변경).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
구조 결함 2개 수정으로 IMP-35 popup 레버를 실동작화:
1. 트리거 모순 해소: 기존 T28.5d 트리거는 eligible_candidate_count==0 요구
— 후보가 '존재만' 해도 미발화인데 그 후보의 자동 소비는 T19d 로 금지
→ 영구 미발화 (mdx04 재현). salvage 정책 차단 터미널
(T19d layout / T28.5 font) 소진 시 overflow zone 을 대상으로 하는
두 번째 트리거 추가 (salvage_exhausted_overflow_pressure).
2. 높이 실감소: slide_base 는 partial 을 그대로 렌더 + <details> 추가일 뿐
(preview_text 는 렌더 경로 미소비 — popup 이 높이를 못 줄였음).
_trim_slot_payload_for_overflow 로 본문 라인을 예산만큼 트리밍 —
재귀 수집 (list-of-str + {text,indent} line-object, f13b 3중 중첩 포함),
라인 경계만, 슬롯당 최소 1라인, label/title 무접촉.
측정 피드백 루프 (최대 3회): 후보 렌더의 잔여 excess 로 추가 트림.
트리밍 전문은 popup body(raw_content 원문)에 보존 — 텍스트 무손실.
실전 검증:
- mdx04: RENDERED_WITH_VISUAL_REGRESSION → **PASS** (trimmed=6, coverage missing 0,
promoted). presentation_ready=False 는 density tight — #15 게이트 정직 판정
- mdx01/02/03/05: 오발화 0, 상태 불변
- SHA baseline 재캡처 (mdx04 final.html 정당 변경), 전체 게이트 1027 passed
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
기계 PASS vs 사람 '빽빽함' 판정 간극을 잡는 결정론적 밀도 신호:
- _estimate_max_card_lines: repeat 슬롯/list-of-dict 컬럼/<br> 라인 payload 추정 (v0)
- _compute_density_gate: px_per_line = (측정 zone 높이 − 56) / 카드 최대 라인
임계값은 mdx 01~05 실측 캘리브레이션 — dense<16 / tight<45 px/line
(mdx03 105.8=ready, mdx02 40~44=cramped, mdx01 bottom 11.5=dense 와 정렬)
- design_readiness 통합: dense/tight → warnings(needs_review) — 사람 판정 정렬
(mdx01 11.5 도 사람 판정은 needs_review 였음 — not_ready 아님)
- presentation_ready 에 density AND 조건 추가 (#98 추천 1: 신호 통합)
- popup 승격 후 재계산 경로에도 동일 배선. 리포팅 전용 — final.html 영향 0
(SHA parity 게이트 무접촉 확인: regression 40/40 포함 전체 1016 passed)
실전 acceptance:
- mdx02: presentation_ready=False + density_tight×2 (44.0/40.4 px/line) — 정직 표시
- mdx03: PASS + comfortable(105.8) 유지, density warnings 0
(mdx03 design_readiness=needs_review 는 기존 reject-label warning — 회귀 아님)
후속: B5 slot metrics(#20) 성숙 시 payload 추정 → DOM 실측 교체
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Gitea #98 Task 12 완결. 롤백된 _emergency_p4_ai_redistribute docstring 이 명시한
승인 아키텍처 5단계('AI: plan only → CODE: apply') 구현.
- _apply_design_adaptation_plan: op 분류 + repeat-count op 만 builder 재실행 적용
- set/increase_repeat_count → _emergency_p4b_build_verbatim_slot_payload
(override_slot_count, repeat 패턴 4종) — max(override, len(groups)) floor 로
원문 그룹 절대 비손실 (Task 11c)
- compact_spacing → policy_skip (T28.5 no_font_shrink) / rebalance_zone_ratio
→ policy_skip (CSS code 소유) / split_group(frame_repeat) → unsupported (#26)
/ frame_contract 변이 → unsupported (catalog 소유) — 전 op 사유와 함께
design_plan_op_status 기록 (침묵 drop 금지, PZ-4)
- count 필드 어휘 관용 추출 (count/repeat_count/to/target_count — 실측 변동)
- prompts.py: task12 계약에 op 스키마 예시 명시
- 시그니처: units 옵션 kwarg (구 호출 backward compat — 미제공 시 명시 status)
검증:
- 신규 테스트 13종 + 영향권/전체 게이트 1003 passed 0 failed
- mdx04 live: 04-1 apply=applied:design_adaptation_plan_no_delta (AI plan 이
apply 계층 최초 통과 — count 5 = 현행 동일이라 no_delta 가 정답), coverage ok
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
11건 전수 분류 결과 프로덕션 결함 0, 전부 '의도된 #98 변경 대비 낡은 테스트':
- A. compute_zone_layout_cols (1건): Task 11d min-width floor(≤260px) 계약으로 갱신
- B. build_layout_css fixtures (3건): floor 반영값으로 재생성 (fixture refresh = 설계된 lock surface)
- C. imp36 self-fire (1건): auto-fit collapsed 0px track 을 점유 트랙 카운트에서 제외
- D. imp88 salvage (5건): T19d(silent layout mutation 금지) 차단 계약으로 재작성
— 차단 터미널/feasibility 선행 차단/전진 금지/trace 필드 4개 측면 + 정책 차단 체인
- E. step17 case_c (1건): 실행 2단 + T28.5 no_font_shrink 정책 차단 터미널로 갱신
- 89a_pre_baseline_sha.json 재캡처 (capture 스크립트 docstring 이 명시한 선행 drift
— #98 스퍼트 기간 누적, 정규 절차)
검증 게이트: pytest tests/phase_z2 + phase_z2_ai_fallback + regression = 990 passed 0 failed
참고: T19d 차단으로 _attempt_salvage_chain 의 layout_adjust 실행 코드는 도달 불가
dead code 상태 — 제거는 #16 (height-fit 재설계) 에서 판단
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
문제: 현재 모든 unit 이 mapper FitError → Emergency P4b verbatim/inline-AI 복구
경로로 흐르는데, 이 경로의 debug_zones 기록이 이미 계산된 placement_trace 를
누락 → PHASE_Z_B4_V4_EVIDENCE=1 에서 v4_evidence_consumed 가 전 zone 관측 불가
(u10 aggregate 테스트 red — 'flag ON이 OFF처럼 동작' silent regression 상태).
수정:
- P4b verbatim(L9107) / inline-AI(L9170) debug_zones 에 placement_trace 추가
(debug 전용 — zones_data/final.html 영향 X. adapter_needed 경로는 기존부터 보유)
- imp95_pre_baseline_sha.json 재캡처 (capture_imp95_pre_baseline.py 정규 절차 —
스냅샷 이후 상류 drift 로 구 baseline 무효화. 89a baseline 과 동일 패턴)
검증:
- trace fields 7/7 + SHA parity 포함 regression 14/14 passed
- tests/phase_z2 전체 709 passed / 11 failed — 실패 11건은 stash 검증으로
본 변경과 무관한 기존 실패 확인 (layout cols/fixtures/imp88 salvage/step17 axis)
참고 (이슈 #10 전제 정정): call-site v4_candidates 배선(L8888)과 Step 11
trace hoist(L8913)는 이미 구현돼 있었음 — 실제 갭은 복구 경로의 trace 유실.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
b836e79 스냅샷에서 UTF-8→CP949 오변환으로 한글 주석·UI 문자열 156줄 손상.
복구 방법:
- CP949 재인코딩→UTF-8 역변환 (ASCII 코드 줄 불변 보장)
- 소실 바이트('?' 잔여)가 있는 줄만 4da22ad 정상본에서 회수 (134줄 + 수작업 9줄)
- 순수 ASCII 줄 변조 0 검증 — b836e79의 기능 변경(pendingSlideBody 등) 전부 보존
- BOM 제거
검증: tsc --noEmit 0 에러, vitest 425/425 통과
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
이슈 #1 점검 결과 figma_to_html_agent는 이미 main에 추적·push되어 있음을 확인
(9fbe3ac 등에서 커밋됨, origin/main에 408 files 존재).
Front_test_v515/ 내 사본은 추적본과 동일 내용의 작업 사본.
남은 갭이던 MCP 의존성 문서화만 반영.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- src: phase_z2 composition/mapper/pipeline/placement_planner/retry, ai_fallback(prompts/schema/validate), mdx_text_atoms 신규
- Front: PipelineTracePanel 신규, FramePanel/SlideCanvas/Home/designAgentApi 등 갱신 + 테스트 4종 추가
- templates/phase_z2: catalog(component_expansion_registry, node_slot_mapping 신규), frames, families, slide_base 갱신
- tests/matching: phase2~26 매칭 실험 스크립트·리포트·온톨로지 전체 (미커밋 진행분)
- tests: b4_v4 evidence, task5~28.5 시리즈, regression(imp95 baseline) 등 신규 테스트 대량 추가
- docs/reference: MDX 구조 인벤토리, MDX→Frame 구조 계약 문서
- scripts: mdx 계약/parity/coverage/viewport 체크, gitea comment, run sync 유틸
- .gitignore: tmp*.json, chromedriver, .orchestrator, *.pkl, Front_test* 등 임시/스냅샷 제외
미완성 작업의 보존용 스냅샷 커밋 (2026-07-02)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Strip the two additive IMP-94 attributes (data-region-id,
data-content-unit-id) symmetrically at both the 89-a fixture capture
script and the b4 mapper source SHA parity test before SHA-256 hashing,
honoring the issue body guardrail "mdx 01-05 의 final.html SHA =
byte-equivalent except for new data-* attrs" without recapturing the
pre-89-a baseline. The strip regex is anchored on the leading-space +
attr-token shape emitted by src/region_marker_stamper.py:131-135 so the
#96 data-frame-slot-id axis stays disjoint.
The marker-parity cross-axis tests for emergency_p4b_verbatim_code and
emergency_p4_ai_inline append sites are converted from pytest.skip to
vacuous-truth early return when the Emergency P4/P4b anchors are absent
in HEAD — the assertion target does not exist in IMP-94 scope, but the
contract still locks placement_markers=[] when the Emergency axis lands
later. Refreshed 89a_pre_baseline_sha.json (2026-05-27T04:19:30Z) holds
the normalized sizes/SHAs for mdx 01-05 post-stamper.
Scope: regression harness + fixture only; zero src/ edits. Verified
35/35 marker-parity + 18/18 SHA parity in a clean detached worktree at
HEAD 2afedfc with these four files applied.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
app_sw_package_vs_solution.html + pre_construction_model_info_stacked.html
were staged as new files but missing from prior commits. catalog
frame_contracts.yaml already references both (family=table / family=list);
this commit reconciles the on-disk partials with the registry so the
13-family baseline matches `git ls-tree` after a clean checkout.
No marker work (data-region-id / data-content-unit-id) — that axis stays
with the marker-injection issue. Disjoint from family/variant
architecture refactor (별 tracking issue).
P7 Patch B used `comments[-1]` at line 1868 but `comments` is defined
inside run_stage, not run_issue scope. The KEEP_OPEN guard runs after
run_stage returns, where `comments` is no longer in scope, causing
NameError crash after Stage 6 YES was already accepted and exit report
generated.
Fix: fetch comments fresh via get_comments(n) at the guard entry.
exit_path file check (fallback) still works as designed.
Refs: #84 (Stage 6 crash during normal close path)
- u1 FramePanel.tsx: extract `applyFrameSelection(candidate, onFrameSelect)`
pure helper; collapse `handleFrameSelect` to direct onFrameSelect for every
V4 label; drop `window.confirm` reject popup (IMP-47B u11 regression noise
per `feedback_auto_pipeline_first`). New vitest pin `imp84_framepanel_reject_silent.test.ts`
covers helper invocation across all 4 V4 labels + source-presence pins.
- u2 templates/phase_z2/slide_base.html: delete `.zone--provisional` CSS,
`.zone__needs-adaptation-badge` CSS, the zone--provisional class fragment
in the zone div, and the badge `<span>` render at the provisional zone.
Preserve `data-provisional="1"` attribute as silent telemetry. New pytest
`tests/phase_z2/test_imp84_provisional_silent_render.py` pins the silent
contract independently of the IMP-30 first-render file.
- u3 tests/test_phase_z2_imp30_first_render.py: invert the three IMP-30 u5
positive provisional-visual assertions to IMP-84 silent-contract negatives
(no class, no badge, no CSS selectors); preserve positive `data-provisional`
telemetry assertions. Docstrings updated to IMP-84 silent contract.
Out of scope (Round #4 + #92 contract): Home.tsx `toast.error(aiReviewMsg)`
call line, designAgentApi.ts `api_error_kinds`/`api_error_kind` schema and
operational-only formatter, FramePanel reject badge/tooltip read-only labels
(L102/L147/L156), and backend `zone.provisional` flag emission.
Stage 4 PASS: u1 vitest 10/10, u2 pytest 5/5, u3 pytest 29/29 (incl. 3
IMP-84 inverted assertions: `test_imp84_provisional_zone_silent_no_class_no_badge`,
`test_imp84_provisional_badge_never_rendered_in_mixed_zones`,
`test_imp84_slide_base_css_strips_provisional_visual_selectors`).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Block Stage 2 YES when IMPLEMENTATION_UNITS contains tests: [].
- Prevent fallback from accepting orchestrator supplement examples as valid plans.
- Honor KEEP_OPEN/DO NOT CLOSE final-close dispositions by skipping close PATCH.
- Add final-close casual self-contradiction guard for YES bodies (allows explicit
`disposition: KEEP_OPEN_*` to pass through to Patch B).
- Inject rejected approaches from failure reports into next-round context with
BANNED_APPROACHES block (tests: [] / DOM mount without jsdom / Home.tsx toast
removal / git add -A).
Refs: #83 (governance break — reopen pending user decision)
#84 (Stage 2 round 5 slip — replay required after this fix)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
u1 KNOWN_AXES tuple gains slide_css entry in src/user_overrides_io.py
(snake_case parity with image_overrides); round-trip test extends
to 6 axes.
u2 src/mdx_normalizer.py surfaces nested slide_overrides.css from the
MDX frontmatter into the normalize_mdx_content return dict; absent
key -> {}, non-string css drops. 4 unit cases in tests/test_mdx_normalizer.py
(present / absent / non-string / title-only).
u3 src/slide_css_injector.py NEW (88 lines) mirrors the
inject_image_overrides_style contract from src/image_id_stamper.py:
marker pair <!--IMP45-SLIDE-CSS:OPEN--> / <!--IMP45-SLIDE-CSS:CLOSE-->,
idempotent re-injection, </head> > <body> > document-start three-tier
fallback, empty/None -> unchanged. 8 fixtures in
tests/test_slide_css_injector.py mirror test_image_id_stamper.py.
u4 run_phase_z2_mvp1 accepts override_slide_css: Optional[str] = None;
None -> frontmatter slide_overrides.css fallback. Step 13 calls
inject_slide_css after image override injection and before the
final.html disk write, so CLI/CI/regression renders observe the same
backend artifact.
u5 argparse adds mutually-exclusive --override-slide-css TEXT (inline
CSS, <style> wrapper optional) and --slide-css-file PATH (UTF-8 read,
fail-closed sys.exit(2) on missing path / decode error / both flags
present). Resolved string is forwarded as override_slide_css kwarg.
6 cases in tests/test_phase_z2_cli_overrides.py (inline / file / both
/ missing / non-utf8 / neither).
u6 samples/mdx_batch/04.mdx frontmatter gains slide_overrides.css
block (verbatim of the former MDX04_DEFAULT_OVERRIDE_CSS constant,
no sample/frame gate). Subprocess smoke in
tests/test_phase_z2_slide_css_smoke.py verifies the marker pair and
CSS substring land in final.html.
u7 Front/client removes the sample/frame-gated frontend-only injection:
Home.tsx drops the MDX04_DEFAULT_OVERRIDE_CSS constant and the
sample==="04"+frame==="process_product_two_way" branch (-28 lines);
SlideCanvas.tsx drops the iframe contentDocument.head injection of
that prop (-14 lines). Live preview now reads backend final.html only.
u8 tests/regression/fixtures/89a_pre_baseline_sha.json 04.mdx entry
resyncs to the live SHA ddb6bf2f... / 28042 bytes (overwrites the
earlier 5-byte-drift d02c76fd... / 28047). Other entries untouched.
Note: 01.mdx baseline drift (ad6f16a3... / 29089 -> live f26a7fac...
/ 29084) predates this branch and is split to a follow-up issue per
the closed-issue fresh validation rule.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PHASE_Z_B4_MAPPER_SOURCE env flag (default OFF) switches slot_payload
source-of-truth from legacy mapper-only / V4 rank-1 to B4 PlacementPlan
.selected_template_id at the single switch site in the runtime loop.
OFF preserves final.html SHA byte-equivalence (u4 parity guard, mdx 01-05).
ON requires Layer A render-active path; BLOCKED exits on B4 no-cover
and on B4-selected FitError (IMP-87 honesty gate pattern — NO silent
fallback). Distinct from PHASE_Z_B4_GATEKEEPER (mismatch render-skip).
Units (1 commit = 1 axis per Stage 1 scope_lock):
u1 — _b4_mapper_source_enabled() flag reader (default OFF)
u2 — _select_mapper_template_id() selector wired at the switch site
u3 — _b4_mapper_source_blocked_exit() for b4_no_cover / b4_selected_fit_error
u4 — render SHA parity regression (tests/regression/ baseline mdx 01-05)
u5 — slot_payload byte-equivalence (matches_mapper=True axis, mdx 01-05)
Targeted 89-a suite 63 PASS; Phase Z regression 323 PASS; IMP-87 mirror
20 PASS. Demo activation via .env only (no vite.config hardcoding).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EMPTY_SHELL_NO_CONTENT overall enum + 3-marker detection (frame_template_id="__empty__"
OR label="empty_shell" OR merge_type="empty_shell") routes empty-placeholder-only
slides to BLOCKED CLI exit 1 + red final_status.html, blocking fake PASS reports
(feedback_artifact_status_naming). Coverage accounting split: legacy covered_section_ids
preserved + new content_rendered_section_ids / empty_shell_section_ids. mdx05 Case B
(zero V4 evidence) honestly classified instead of synthesizing fabricated rank-1 reject
frames. IMP-30 u6/u7 stale empty-shell PASS assertions inverted (29 tests). IMP-85 smoke
parametrize: mdx05 removed from exit-0 list + dedicated BLOCKED exit test added (4 tests).
No production behavior change for chain_exhausted Case A; no AI route activation; no
mdx-id hardcoding. 53 targeted + 76 adjacent Phase Z tests PASS.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a test-only invariance gate that locks the pre-existing four-test red
baseline so IMP-35 cannot silently grow the red surface while in-flight.
u11 does NOT fix the four reds — Stage 2 follow_up_candidates tracks the
actual repair as a separate issue. u1~u10 production work remains in the
worktree and is explicitly out of this commit per Stage 3 R7 carve-out.
Frozen registry (IMP35_BASELINE_RED_NODE_IDS, set semantics):
1. tests/test_imp47b_step12_ai_wiring.py
::test_mixed_units_classified_by_route_and_provisional_flag
2. tests/test_imp47b_step12_ai_wiring.py
::test_reject_provisional_unit_reaches_router_short_circuit
3. tests/test_imp47b_step12_ai_wiring.py
::test_step12_ai_repair_artifact_writes_json_serialisable_records
4. tests/test_phase_z2_ai_fallback_config.py
::test_ai_fallback_master_flag_default_off
Gate semantics (subprocess pytest, set comparison):
- All 4 node ids resolve to collectible pytest items
(rename / delete is caught up front).
- Broader baseline-area sweep across the two registry files yields
EXACTLY 4 FAILED and 0 ERROR, with FAILED set ≡ registry.
- A new red in the baseline area flips count above 4 OR introduces a
FAILED id outside the registry; either branch fails the gate.
- Cross-lock test ensures registry node ids cannot point outside the
declared area-files inventory.
AI isolation contract (feedback_ai_isolation_contract):
Gate body uses stdlib only (subprocess + re + ast). An AST self-verify
test rejects `anthropic` imports and `route_ai_fallback` references in
this file, structurally preventing AI routing inside the gate.
Stage 4 verification (HEAD c1df656 pre-commit):
pytest -q tests/phase_z2/test_imp35_baseline_red_invariance.py
→ 7 passed in 15.26s.
Baseline area sweep
(tests/test_imp47b_step12_ai_wiring.py +
tests/test_phase_z2_ai_fallback_config.py)
→ 4 failed / 6 passed / 0 errors; FAILED set ≡ registry (identity).
pytest --collect-only on the 4 registered node ids → all 4 resolve.
py_compile clean. Codex R1 = YES (independent verify).
Guardrails honored:
- Scope-locked: test-only file; zero production code in this commit.
- 1 commit = 1 decision unit (u11 only).
- No hardcoding: registry = Stage 2 contract frozen tuple, not
sample-specific literal; gate body has zero magic constants.
- AI isolation: stdlib-only gate, AST self-verify locks isolation.
- baseline-red 4 body repair = separate follow-up issue, not u11 scope.
source_comment_ids: Stage 1 problem-review; Stage 2 plan R2 + Codex R2
YES; Stage 3 Claude #30 + Codex #31 R7 YES; Stage 4 Claude #32 + Codex
#33 R1 YES.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add resplit_all_reject_merges() helper in phase_z2_composition.py that
detects parent_merged / parent_merged_inferred units with label=reject
and rebuilds them as per-section single units using each section's own
rank-1 V4 evidence (no frame swap, MDX raw_content preserved).
Pipeline hook fires once after Step 6 settling chain (u12/u4/empty-shell)
and section_assignment_plan resolution, before Step 6 artifact write.
Guards: beneficial-split rule (>=1 non-reject), coverage equality, layout
cap (>4 abort), max_retry=1, section_assignment_override short-circuit.
Audit: comp_debug["imp48_resplit"] additive payload (applied, split_units,
skipped_units, post_split_unit_count, post_split_layout_preset);
selection_path="resplit_from_merge" telemetry on rebuilt singles;
layout_preset re-derived via select_layout_preset(new_units).
Tests: 39/39 PASS (composition u1~u6: 14 cases; pipeline u7~u9: 25 cases).
Scoped regression 720/6 with 6 failures isolated as pre-existing on
baseline 79f9ea5 (independent of IMP-48). mdx03 golden lock preserved.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
