Transcript-first, lazy-visual-verification video understanding toolkit
(va CLI + agent skill).

Understand long videos without watching every frame: whisper transcript
-> hypotheses -> frame verification only where uncertainty lives -> an
append-only checkpoint ledger you can query, clip, and hand off.

Two loops, two ledgers: the understanding loop (Kubrick engine) writes
facts; the editing loop (Kuleshov) writes cut decisions and can never
mutate facts. EDL / FCP7 XML / FCPXML / OTIO / SRT handoff carries
grounding labels, revision pins, and a deterministic Receipt sidecar,
with NTSC-exact rational timecode.

Highlights of this release cycle:
- Live demo GIF in the README, recorded from the real corpus browser
- Responsive corpus browser with shareable URL state and #cp deep links
- Incremental corpus projections (fingerprint-skip unchanged work)
- Workspace audio cache shared by ingest / diarization / audio events
- Parallel frame capture (3.65x measured) with process-group-safe
  download cancellation
- Transcript-collapse guard against silent long-form decode stalls
- Speaker diarization as the one optional extra ([diarize]); ungated
  sherpa backend works out of the box
- Rational stream-timing metadata persisted end to end
- Fully documented CLI surface with consistent Korean error guidance

Install: clone + 'uv tool install --python 3.12 .[diarize]'
(or scripts/install.sh). Requires ffmpeg/ffprobe; yt-dlp for URL ingest.
macOS and Linux supported; Windows experimental (CI-smoked).

Benchmarks reproduce from CC-BY Blender open-movie fixtures:
uv run python benchmarks/run_bench.py --set public --fetch