One ffmpeg Blind Spot, Three Repos: The EOF-Freeze Fixes

Yesterday I shipped AIVideoSentry v1.1.0 with a fix for a nasty ffmpeg quirk: freezedetect never prints freeze_end — or freeze_duration — when a freeze runs to end of file. The output is just an orphan freeze_start with nothing after it. Any code that trusts the end marker silently passes the most common stall shape in AI video: the generator dies, the last good frame clones to EOF, the file looks 8 seconds long and the tail is dead.

The uncomfortable question after that fix: how many of my other repos trust the same marker?

The audit

A grep across the safety-net series gave two more hits:

• AIVideoQualityGate — its freeze regex required freeze_start and freeze_end on the run. Root cause confirmed with a real fixture: a render with a 5-second frozen tail passed QC with exit 0.
• ComfyUI-VideoGate — its parser read the JSON event lines but required the freeze_duration key. Same fixture: zero issues reported on a half-dead render.

Same defect, two more repos. Same 5-second proof I'd built for Sentry.

The fixes

Both repos now reconstruct the orphan event from freeze_start + the real video duration:

• QualityGate prefers frame math (nb_read_frames / avg_frame_rate — real frames, immune to short audio tracks that shrink container duration) and falls back to stream then format duration.
• ComfyUI-VideoGate pairs the orphan freeze_start with probe_result.duration_s, wired through engine.py, and labels the event honestly: frozen frames 5.04s at t=2.96s (ran to end of file).

And the audit caught something my fixtures didn't. Reviewing the VideoGate parser, I found a second defect hiding behind the first: freezedetect emits each event as separate single-key lines — freeze_start, freeze_duration, freeze_end each arrive on their own line — so the same-record parser was also dropping real mid-stream freezes. The fix became a proper pairing state machine: complete triples pair up as normal events, lone starts reconstruct against EOF. 40 tests green, including a new mid-stream pairing test built for exactly that bug.

What a paid model caught

I ran both diffs through a fresh-eyes review on a frontier model. On QualityGate it earned its keep — four cheap, real hardenings I applied the same day:

1. Prefer avg_frame_rate over nominal r_frame_rate in the frame math (VFR/telecined content could badly skew the reconstructed end time).
2. Skip attached_pic cover-art streams — a 1-frame still's duration would truncate the reconstruction.
3. Never reconstruct from a failed ffmpeg run — an interrupted decode can leave an unmatched freeze_start that isn't an EOF freeze at all. Now it returns "unavailable" instead.
4. And its reminder about container-vs-video duration is folded into the frame-math-first ordering.

The honest caveat: through this particular relay the premium reasoning models streamed zero visible tokens (empty replies after thousands of completion tokens — a tooling limitation, not a model one). The working reviewer was the fast flagship instead, and it did genuine work. Credit notes are nice; verification is better: every fix above is proven by fixtures, exits, and suites, not by vibes.

The pattern worth stealing

If you call freezedetect anywhere: grep your parser for freeze_end and ask what happens when it never comes. The EOF-orphan shape isn't an edge case in generative video — it's the dominant failure mode. A stalled generator doesn't produce corrupt files. It produces valid files with dead tails.

Final state: ComfyUI-VideoGate v1.1.0 (40/40 tests), QualityGate patched on main (5/5), Sentry patched earlier in v1.1.0 (25/25). Three repos, one ffmpeg line each, and the tails are dead no more.