Handover: where this session stopped

The overnight extraction is in flight with its paths and its restart
instructions. Repository state for all four toolboxes. tracks-wip is exact now
but should not merge until ARJ decides whether MGT wants a second extractor at
all, and its videogram pyramid has never been run.

Ends on the tally of my own faults today, because three of them were one fault in
different clothes: a test reaching for the convenient fixture rather than the one
that reproduces the bug. Running it against the broken code is the step that
caught all three.

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

@@ -230,3 +230,87 @@ 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.
+
+---
+
+# Where this session stopped
+
+## In flight: the overnight extraction
+
+**Running now**, started 2026-08-24 late evening, expected about 1.4 h.
+
+    /tmp/claude-1000/run_qom.py          the script, self-documenting
+    /tmp/claude-1000/qom_night.log       its output
+    .../PanasonicDownsized/analysis/27CoLocated.Panasonic.A003C505_231127_DJ0B/
+        qom.csv        the motion envelope, one row per frame
+        qom_run.json   the command, the timings, and what QomRaw is
+
+`mg_motion(motion_analysis='qom', save_motiongrams=False, normalize=False)` over the whole
+session at native 1920x1080. The **validated** path, not `_tracks.py`.
+
+*If it is not there when you read this*, check the log first: this run already died once,
+silently, because MGT could not open the file at all (see the framecount fault above). The
+script is idempotent --- just run it again.
+
+*What it gives you:* the envelope every later step needs --- the hierarchy, the
+action/phrase/part levels, and the audio-video timing. Not the videogram pyramid; that
+needs `_tracks`.
+
+## Repository state
+
+| repo | branch | state |
+|---|---|---|
+| ambiscape | main | clean, in sync, 0.45.0 |
+| musiscape | main | clean, in sync, 0.8.0 |
+| micromotion | main | clean, in sync, 1.15.0 released |
+| MGT-python | master | clean, in sync, **1.14.1 released and on PyPI** |
+| MGT-python | tracks-wip | clean, in sync, **not merged** |
+
+`master` carries the quadratic fix (released) and the framecount fix (unreleased).
+
+## `tracks-wip`: ready for review, not for merge
+
+The extractor is now **exact** --- byte-identical to `mg_motion`'s QomRaw across 6,005
+frames of 1080p, serially and in parallel. 120 s of video in 60 s serial, 47 s over eight
+workers; the full session about 1.0--1.3 h against `mg_motion`'s 1.4.
+
+Before it merges, three things:
+
+1. **Decide whether MGT should have it at all.** It duplicates what `mg_motion` does, for
+   the case `mg_motion` is not built for. The alternative is a `lean=True` path inside
+   `mg_motion` rather than a second module.
+2. **It is not exported** from `__init__.py`, deliberately.
+3. **The videogram pyramid is written but unexercised.** `build_pyramid` and `read_columns`
+   have no tests at all --- they were designed and coded, never run against real material.
+
+## The honest tally of my own faults today
+
+Written down because the pattern is worth more than the individual bugs, and because three
+of them were the same fault in different clothes.
+
+- **Three tests that could not fail.** A timing test for the quadratic bug that passed
+  against the unfixed code; a seam test written twice, with two different fixtures, that
+  passed against the broken code both times; and a strict `xfail` that XPASSed. In every
+  case I reached for the convenient fixture rather than the one that reproduces the fault,
+  and only found out by running against the broken code. **That step is the one to keep.**
+- **A monitor that watched itself.** `pgrep -f "run_qom.py"` matches the watcher's own
+  command line, so it waited forever on a process that had died 24 minutes earlier.
+- **A patch applied to the wrong function.** The `Popen`/`communicate` block appears several
+  times in `_utils.py`; a single-occurrence replace landed in `ffprobe()`.
+- **A changelog entry that was never written**, while the script printed "changelog updated"
+  unconditionally --- a success message not conditional on success.
+- **A blocker I asserted twice and was wrong about both times**, in the Sound Spaces sweep:
+  claiming a report could not be re-measured because a deposit was unpublished, when the
+  deposit tree was local and only a *reader* was blocked.
+
+## What is waiting on you
+
+**On the dance work**, in the design doc's own words: where the cached table lives, which
+level the student annotates first, and the round-trip question. Plus, new from the probe:
+whether MGT should carry `_tracks` at all.
+
+**On MGT**, unchanged: #373's three Laban questions, #359's dependency, #312's 2023
+proposal, #311's author list, and 2.0 waiting one more release.
+
+**On Sound Spaces**, unchanged: publish StillStanding365, the loudest-room rule, the audio
+crossing fraction, and where shared report code lives.