Fix slide-transcript mismatch: add frame extraction to pipeline, skip screenshare videos in alignment

Three fixes for the screenshare/slide mismatch:

1. Pipeline now runs frame_extractor.py between Transcribe and Align.
   Previously frame extraction was never called, so screenshare (SS)
   videos always fell through to PDF-based slide matching.

2. semantic_alignment.py now checks each caption's existing alignment
   before attempting slide matching. If the alignment has
   source="screenshare" (produced by frame_extractor), the caption is
   skipped — preventing PDF matching from overwriting frame alignment.

3. frame_extractor.py skip logic now checks if existing alignment is
   frame-based before skipping. If a slide-based alignment exists but
   the video is a screen recording, frame extraction still runs and
   produces the correct frame-based alignment.

Pipeline flow is now:
  Download → Transcribe → Extract Frames → Align → Generate Notes

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: nodeeeeee

electron/renderer/app.js

@@ -1495,6 +1495,8 @@ async function attachPageHandlers() {
         chain.push(['Transcribe', c]);
       }
       if (steps.includes('align')) {
+        // Extract frames from screenshare videos before slide-based alignment
+        chain.push(['Extract frames', [python, paths.frame_extractor, '--course', cid, '--path', outDir]]);
         const c = [python, paths.align, '--course', cid];
         if (force) c.push('--force');
         chain.push(['Align', c]);

frame_extractor.py

@@ -632,11 +632,17 @@ def process_course(course_id: str, base_dir: Path) -> int:
         frame_dir = course_dir / "frames" / video_stem
         align_file = course_dir / "alignment" / f"{video_stem}.json"
 
-        # Skip if already processed
+        # Skip if already has frame-based alignment
         if align_file.exists():
-            print(f"  [skip] Already extracted: {video_stem}")
-            processed += 1
-            continue
+            try:
+                with open(align_file, encoding="utf-8") as _af:
+                    if json.load(_af).get("source") == "screenshare":
+                        print(f"  [skip] Already extracted: {video_stem}")
+                        processed += 1
+                        continue
+                # Existing alignment is slide-based — re-extract frames
+            except Exception:
+                pass
 
         # Find matching caption
         caption_path = course_dir / "captions" / f"{video_stem}.json"

semantic_alignment.py

@@ -1684,6 +1684,20 @@ def process_course(course_id: int | str, use_jina: bool = False,
         except Exception:
             pass   # can't read → proceed and let align() handle it
 
+        # Skip captions that already have frame-based (screenshare) alignment.
+        # Those were produced by frame_extractor.py and should not be overwritten
+        # with slide-based alignment.
+        _existing_align = out_dir / f"{cap.stem}.json"
+        if _existing_align.exists():
+            try:
+                with open(_existing_align, encoding="utf-8") as _af:
+                    _align_data = json.load(_af)
+                if _align_data.get("source") == "screenshare":
+                    print(f"  [skip] Screenshare alignment exists: {cap.stem} (use frames)")
+                    continue
+            except Exception:
+                pass
+
         # ── Priority 1: user-supplied mapping ────────────────────────────────
         slide_group: list[Path] = []
         if cap.stem in user_mapping: