fix(#75): IMP-47A mdx03 frontend execution stabilization (u1~u4)

u1: SlideCanvas iframe sandbox += allow-scripts (allow-same-origin preserved)
    → embedded-mode script in slide_base.html now applies html.embedded
    → standalone CSS reset deactivates inside iframe; no clipping
u2: designAgentApi.loadRun merges candidate_evidence + v4_all_judgments
    + v4_candidates via Map<template_id|id|frame_id> dedup,
    LABEL_PRIORITY (use_as_is<light_edit<restructure<reject) then
    confidence desc, capped TOP_N_FRAMES=6
u3: Home.handleGenerate useCallback deps = [uploadedFile, slidePlan,
    userSelection, pendingZones, pendingLayout] (5-tuple, stale-closure fix)
u4: tests/manual/imp47a_e2e.md — mdx03 manual e2e spec (5 axes)

Frontend-only. Backend src/ untouched. No template/catalog edits.
Determinism preserved (no LLM in frontend merge logic).
Baseline: pytest -q tests → 623 passed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Front/client/src/components/SlideCanvas.tsx

@@ -293,7 +293,7 @@ export default function SlideCanvas({
               title="Phase Z 렌더 결과"
               className="w-full h-full border-0 block"
               scrolling="no"
-              sandbox="allow-same-origin"
+              sandbox="allow-same-origin allow-scripts"
               style={{ pointerEvents: isEditMode ? "auto" : "none" }}
               onLoad={(e) => {
                 // IMP-14 (Step 13 A-4) — embedded vs standalone CSS reset 은 backend

Front/client/src/pages/Home.tsx

@@ -377,7 +377,7 @@ export default function Home() {
       );
       setState((p) => ({ ...p, isLoading: false }));
     }
-  }, [state.uploadedFile]);
+  }, [state.uploadedFile, state.slidePlan, state.userSelection, pendingZones, pendingLayout]);
 
   // ── 섹션 드래그 앤 드롭 (Zone으로 재배치) ──
   const handleSectionDrop = useCallback((sectionId: string, zoneId: string) => {

Front/client/src/services/designAgentApi.ts

@@ -511,12 +511,17 @@ export async function loadRun(runId: string): Promise<LoadRunResult> {
     const candidateEvidence = Array.isArray(unit.candidate_evidence)
       ? unit.candidate_evidence
       : [];
-    const rawSource =
-      candidateEvidence.length > 0
-        ? candidateEvidence
-        : (unit.v4_all_judgments?.length > 0
-            ? unit.v4_all_judgments
-            : (unit.v4_candidates ?? []));
+    const candidateMap = new Map<string, any>();
+    const pushCandidate = (c: any) => {
+      if (!c) return;
+      const key = c.template_id ?? c.id ?? c.frame_id;
+      if (!key) return;
+      if (!candidateMap.has(key)) candidateMap.set(key, c);
+    };
+    candidateEvidence.forEach(pushCandidate);
+    (unit.v4_all_judgments ?? []).forEach(pushCandidate);
+    (unit.v4_candidates ?? []).forEach(pushCandidate);
+    const rawSource = Array.from(candidateMap.values());
     const v4Source = [...rawSource].sort((a: any, b: any) => {
       const lp = (LABEL_PRIORITY[a.label] ?? 99) - (LABEL_PRIORITY[b.label] ?? 99);
       if (lp !== 0) return lp;

tests/manual/imp47a_e2e.md

@@ -0,0 +1,90 @@
+# IMP-47A — mdx03 frontend stabilization manual e2e
+
+Scope: frontend-only. Backend pipeline must NOT be modified during this test.
+Path under test: `mdx=03` (default sample loaded on page open).
+
+## Preconditions
+
+- Backend running on `http://localhost:8001` (`uvicorn src.main:app --port 8001`).
+- Frontend dev server running (`cd Front && npm run dev`).
+- Working tree at IMP-47A Stage 3 HEAD (u1+u2+u3 applied to `Front/client/src/components/SlideCanvas.tsx`, `Front/client/src/services/designAgentApi.ts`, `Front/client/src/pages/Home.tsx`).
+- Browser opens `http://localhost:5173/?mdx=03` (or default route, which auto-loads `mdx=03`).
+
+## Section 1 — iframe rendering (axis 1)
+
+Goal: verify u1 sandbox change lets `slide_base.html` script apply `html.embedded` class so the slide is not clipped inside the iframe.
+
+Steps:
+1. Open the app; wait for `mdx=03` auto-load and initial `final.html` render.
+2. Click the "슬라이드 플랜 생성하기" button; wait for `run "<id>" 완료` toast.
+3. Open browser DevTools → Elements; locate the iframe inside `SlideCanvas`; switch context to the iframe document.
+4. Confirm `<html class="embedded">` is present (not just `<html>`).
+5. Confirm the rendered slide content fills the 1280×720 frame with no top padding offset and no clipping at the bottom.
+
+Pass: `html.embedded` class present AND no visible vertical shift/clipping.
+Fail signal: iframe content pushed downward, footer cut off, or `html` lacks `embedded` class (means script never ran → sandbox regression).
+
+## Section 2 — multi-source frame candidates (axis 2)
+
+Goal: verify u2 3-source merge surfaces candidates from `candidate_evidence`, `v4_all_judgments`, and `v4_candidates` with deterministic dedup and cap.
+
+Steps:
+1. After Section 1 success, click any zone in the canvas; right panel switches to the "frame" tab.
+2. In the frame candidate list, count visible candidates. Expect ≤ `TOP_N_FRAMES` (=6).
+3. Open DevTools → Network → reload `/api/run/<id>`; inspect the JSON response and confirm at least two of `candidate_evidence`, `v4_all_judgments`, `v4_candidates` are non-empty.
+4. Cross-check: union of `template_id ?? id ?? frame_id` keys from all three arrays (deduped, capped at 6) equals the UI list count and order.
+5. Confirm the order respects LABEL_PRIORITY (`use_as_is` < `light_edit` < `restructure` < `reject`) then descending confidence.
+
+Pass: union/dedup/cap/order all match.
+Fail signal: only candidates from a single source visible, duplicates by template_id, more than 6 items, or order violates LABEL_PRIORITY.
+
+## Section 3 — frame / layout override regeneration (axis 3)
+
+Goal: verify u3 5-dep `handleGenerate` callback delivers the latest override state to backend (no stale closure).
+
+Steps:
+1. From Section 2, pick a non-default frame candidate (one whose label is not `use_as_is`); click "이 프레임 적용".
+2. Confirm bottom-left button transforms to "선택대로 재생성하기" with amber pulse dot (hasPendingChanges = true).
+3. Without any extra clicks, click "선택대로 재생성하기".
+4. Open DevTools → Network → inspect the POST `/api/pipeline` body; confirm `overrides.frames` contains the chosen `{unit_id: frame_id}` mapping.
+5. After success toast, confirm new `run_id` differs from the previous run, and the rendered iframe reflects the chosen frame (frame DOM class / id matches selection).
+6. Repeat with a layout-card "적용하기" → pending overlay enters → "선택대로 재생성하기"; confirm POST body includes `overrides.layout` with the chosen preset id.
+
+Pass: every override (frames / layout / zoneSections / zoneGeometries when applicable) reaches backend on first click; new `run_id` returned.
+Fail signal: POST body lacks the override, or backend re-renders with the previous selection (stale closure regression).
+
+## Section 4 — pending overlay enter / cancel / clear (axis 4)
+
+Goal: verify pendingLayout overlay enters on "적용하기", exits on "취소", and auto-clears on successful regenerate.
+
+Steps:
+1. Click any non-current layout card "적용하기" button.
+2. Confirm amber dashed overlay appears over `.slide-body` area with `PENDING BODY LAYOUT` label and chosen layout id.
+3. Confirm "취소" button (top-right of canvas) is visible.
+4. Click "취소"; confirm overlay disappears, `userSelection` resets, and `hasPendingChanges` indicator clears.
+5. Re-enter pending mode (apply a layout again), this time click "선택대로 재생성하기"; confirm overlay disappears on success (Home.tsx clears `pendingLayout` + `hasPendingChanges` before pipeline call) and the new `final.html` renders inline (no overlay).
+
+Pass: enter → cancel → re-enter → regenerate cycle leaves no overlay, no stuck pending state, no leftover `hasPendingChanges` flag.
+Fail signal: overlay persists after regenerate, button stays in amber state, or cancel does not restore the canvas.
+
+## Section 5 — mdx03 end-to-end pass (axis 5)
+
+Goal: smoke run combining Sections 1–4 in one session to validate mdx03 demo path.
+
+Steps:
+1. Fresh reload `http://localhost:5173/?mdx=03`.
+2. Click "슬라이드 플랜 생성하기" → wait for run completion → confirm iframe renders cleanly (Section 1 pass).
+3. Click any zone → inspect 2+ candidates in frame list (Section 2 pass).
+4. Apply a non-default frame → "선택대로 재생성하기" → new run id + iframe reflects override (Section 3 pass).
+5. Apply a non-default layout → confirm overlay → "선택대로 재생성하기" → overlay clears (Section 4 pass).
+6. Capture: run_id chain, final.html path under `data/runs/<run_id>/final.html`, and DOM screenshot of the rendered iframe content.
+
+Pass: all five sections green, no console errors, no toast errors, three distinct `run_id`s produced across the session.
+Fail signal: any earlier section regression OR backend pipeline failure (out-of-scope for IMP-47A — log separately).
+
+## Out of scope (not tested here)
+
+- AI fallback activation in `Step 12` (`light_edit` / `restructure`) — IMP-47B.
+- Frame cache (#62 / IMP-46).
+- mdx04 / mdx05 path-specific axes.
+- Automated Playwright e2e replacement — future work.