Handover: the framecount ceiling, 1.14.1, and the tracks resolution

A second fault of the same shape as the quadratic one: get_framecount's flat 10 s
timeout meant MGT could not OPEN a video longer than about twenty minutes, and
the fallback escalated to a slower method so it could never recover. Found
because the overnight run died instantly.

Both bugs were invisible on a clip and fatal on a session, and neither changed a
number. MGT's suite is entirely short-clip, so nothing in it could have found
either --- which is the argument for the property tests added rather than for a
two-hour fixture in CI.

Also records 1.14.1, the two slips in cutting it, and that _tracks' non-determinism
was two output specs feeding one stdout.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xg1f739wddu5M4s3UwdNkn

Author: alexarje

plans/2026-08-24-handover.md

@@ -184,3 +184,49 @@ it recording the command and the elapsed time. About 1.4 h expected.
 
 That gives the motion envelope the hierarchy and the audio-video timing both need. The
 videogram pyramid waits for `_tracks` to have tests that catch what its author did not.
+
+## And a second fault of the same shape: long videos could not be OPENED
+
+Found because the overnight run died instantly. `get_framecount` gave ffprobe a flat
+**10-second timeout** and, on timeout, escalated to `-count_frames` --- which fully decodes
+and is far slower --- so that timed out too and raised. A ceiling on video length with
+nothing to do with any analysis.
+
+Counting packets on the 2 h 38 min recording takes **22 s**, so `MgVideo()` raised
+*"Could not count frames. (Is this a video file?)"* on an ordinary MP4. Anything past
+roughly twenty minutes was unopenable. It opens in 21 s now and reports 475,680 frames.
+
+The allowance scales with the file (`framecount_timeout()`, separate so a test can shorten
+it and exercise the real path against a real ffprobe), and a timeout falls back to the
+container's `nb_frames` with a `RuntimeWarning` instead of escalating. Four tests, which
+error on the old code.
+
+**These two bugs share a shape and that is the lesson.** Both were invisible on a clip and
+fatal on a session, and neither changed a number. MGT's suite is entirely short-clip, so
+nothing in it could have found either. The property tests added here --- count the
+allocations, shorten the timeout --- are the cheap way to cover that class without a
+two-hour fixture in CI.
+
+## Released 1.14.1
+
+The quadratic fix, byte-identical, on PyPI. CI green on the release commit *before* the
+Release was created, which is the rule this handover records being broken earlier the same
+day. The framecount fix is on master and unreleased.
+
+*Two slips of my own in that release, both in the commit record.* The changelog entry the
+fix commit claimed to add was never added: a `str.replace` targeted an `## [Unreleased]`
+heading that 1.14.0 had already consumed, matched nothing, and the script printed "changelog
+updated" unconditionally --- a success message not conditional on success. And I nearly cut
+the release from the wrong branch, because the notes extraction kept returning nothing while
+I was still on `tracks-wip`.
+
+## `_tracks.py`: the non-determinism is solved
+
+`ffmpeg_cmd(pipe="read")` appends its OWN output arguments, and the module had appended its
+own. ffmpeg was given two outputs and wrote both into one stdout, interleaved --- wrong
+values, and different between runs because interleaving follows buffering. Serial extraction
+is now byte-for-byte equal to `mg_motion` over 6,005 frames of 1080p.
+
+**The parallel path is still not exact**: one repeated frame at the last chunk seam, because
+`-ss` before `-i` lands on a keyframe and the frames decoded before the target are not always
+the single one the worker drops. Seven interior seams are exact. Do not use it until that is.