refactor: decouple open_file from recording logic + auto-detect ffmpeg

[sonichi]

Summary

Per owner's design directive: recording tools return explicit file paths, open_file becomes a path-based file opener, and findFfmpegWithSubtitles() auto-detects the right ffmpeg binary for subtitle burns.

Supersedes #355 (subtitled_pending gate) and #357 (ffmpeg-full path switch) — folds both concerns into a cleaner architecture.

Changes (+71/-41)

1. open_file stripped of recording-specific logic

• Requires path arg — get it from the recording tool's return value
• find_recording: true flag as explicit fallback only (when model lost the path)
• Removed: subtitled_pending flag, version detection, findRecording() as default behavior
• Kept: known file aliases (diagnostics), QuickTime activation, duration/size metadata

2. screen_record stop returns explicit file list

• New files object: {raw, narrated, subtitled, recommended}
• instruction string tells the model exactly which path to pass to open_file
• Subtitle burn still happens synchronously on stop

3. findFfmpegWithSubtitles() auto-detection (from #357)

• Probes: $FFMPEG_SUBTITLE_BIN env → system ffmpeg → homebrew narrow → homebrew full (keg-only)
• Cached per process lifetime (~50ms first call)
• burnLiveTranscriptSubtitles() uses it instead of hardcoded /opt/homebrew/bin/ffmpeg
• If no binary has libass, logs clearly and skips burn (no cryptic AVFilterGraph error)

Why this design

Owner: "The recording tool should return the files and the default file to use. The caller should process the result and report abnormality. Then open_file should be called with filename."

This eliminates the class of bugs we chased yesterday:

• No more subtitled_pending false positives (flag is gone)
• No more 18s polling timeout (loop is gone)
• No more "which ffmpeg" confusion (auto-detected)
• Model gets explicit file paths and instructions, not implicit guessing

Test plan

• npx tsc --noEmit --skipLibCheck clean
• findFfmpegWithSubtitles() runtime-tested: correctly picks ffmpeg-full on Mac Mini
• Manual: phone call "record 20s then open it" — model should pass the path from screen_record result
• Manual: verify subtitle burn uses the detected ffmpeg-full binary

PRs to close after merge

• fix: subtitled_pending false positive in open_file #355 — subtitled_pending gate (superseded by removing the flag entirely)
• subtitle burn: use ffmpeg-full for libass-dependent filter #357 — ffmpeg-full path switch (folded into this PR's findFfmpegWithSubtitles)

Still open (independent)

• retention: archive stale results/*.txt on startup (D half of flood fix) #354 — retention sweep D (results/ cleanup, separate concern)

🤖 Generated with Claude Code

[sonichi]

MacBook review: LGTM. Clean redesign:

• open_file now requires path from model, find_recording is explicit fallback only
• screen_record stop returns {files: {raw, narrated, subtitled, recommended}} with paths
• findFfmpegWithSubtitles() cached probe with env override
• burnLiveTranscriptSubtitles uses detected binary, graceful skip if no libass
• Removed all subtitled_pending/polling logic (root cause of false positives)
• +71/-41, tsc clean. Ready to merge.