animation: inkstats undercounts the final hold of a rotoscope work dir

[kyle-sexton]

For a rotoscope work directory, inkstats.drawings() ends the last drawing one nominal frame after its start instead of at d/index.json's duration, so the final hold is undercounted (regression clip: 0.042 s counted, 0.133 s actual). Found by Codex on #4534; deferred there because the fix changes the learned woodcut-ink pack.

Fix (tested, not applied)

diff --git a/plugins/animation/scripts/inkstats.py b/plugins/animation/scripts/inkstats.py
index 06465ed4b..c1d15a83d 100755
--- a/plugins/animation/scripts/inkstats.py
+++ b/plugins/animation/scripts/inkstats.py
@@ -120,8 +120,13 @@ def drawings(film, fps, region=None, window=None):
             continue
         prev = rgb.astype(np.int16)
         yield rgb, t
-    else:   # the film ended: a video where its stream ends, anything else one frame after the last
-        t = decode.end(film) if isinstance(film, (str, Path)) and Path(film).is_file() else t + dt
+    else:   # the film ended: a video where its stream ends, a work dir at its index's duration, else one frame on
+        if isinstance(film, (str, Path)) and Path(film).is_file():
+            t = decode.end(film)
+        elif isinstance(film, (str, Path)) and decode.workdir.src(Path(film)).is_dir():
+            t = json.load(open(decode.workdir.index(Path(film)), encoding='utf-8'))['duration']
+        else:
+            t += dt
     yield None, t
 
 

Result of relearning the pack with the fix

• controls.py check: 36/36 as required against both the old and the relearned pack; no control flips to passing. Selftest unchanged.
• Held-out: calibration 152/152, evaluation 147/150 (was 150/150). Splits 151 (76 calibration) instead of 150.
• offstep: band 0.0308-0.1406 (source 0.0882) becomes 0.0393-0.1242 (source 0.084), widening 0.0096 to 0.0007; the 3 held-out misses are all on this row.
• Also moved: per_second 7.784-8.04 to 7.777-7.968 (source 7.919 to 7.895); flat 0.1305-0.7958 to 0.1057-0.819; duration 30.182 to 30.273.

To decide

Accept 147/150, or change how bands and widening are learned so held-out coverage survives the corrected statistic, then re-baseline the pack and the refactor safety net.

Related

• feat(animation): animation plugin with rotoscope, learn-style and setup skills #4534 (Codex thread on plugins/animation/scripts/inkstats.py)
• animation: woodcut-ink check residuals after scene rounds 7-10 #4507 (woodcut-ink check residuals)

[kyle-sexton]

This was generated by AI during triage.

Triage: verified bug, human-gated on the pack decision

Verified (code read on main). The drawing iterator's end-of-film branch only special-cases a video file. A rotoscope work dir is a directory, so it falls to "one nominal frame after the last drawing's time" and ignores the duration recorded in the work dir's drawing index. That explains the reported shortfall: the final hold is counted as 1/fps, not as the time left until the clip ends. The frame decoder already reads that same index for a work dir, so the fix in the body uses a value that is already there.

Brief

• Behavior: for a work-dir input, the last drawing's hold ends at the index's duration. Video-file inputs, frame folders, and synthetic iterables keep their current end rules.
• Acceptance: the regression clip counts 0.133 s for its final hold, not 0.042 s. controls.py check stays 36/36 and the selftest is unchanged. The woodcut-ink pack is relearned and committed together with the fix, and held-out coverage meets whatever bar is decided below.
• Why a human decides: the fix is small and already tested. The open question is the pack. With the corrected statistic, held-out evaluation drops from 150/150 to 147/150, and all three misses are on offstep. There are two choices: (a) accept 147/150 and re-baseline, or (b) change how bands and widening are learned so held-out coverage survives, then re-baseline the pack and the refactor safety net. That is a calibration-method choice with no single right answer, so it is not defaulted.
• Once decided: (a) is scoped work: apply the diff, relearn, commit the pack. (b) first needs a design for the band-learning change.

Type: Bug. Priority: medium (nothing is blocked, and the error is a bounded undercount on one statistic).

[kyle-sexton]

This was generated by AI during work-loop.

Ratify C3 dispatch? This item now carries work-class: scoped and agent-ready (applied 2026-09-27 18:52Z), and the work-loop admission gate (instance melo-lap-001-wsl-2) reads it as C3, bug-fix-shaped. It has no recorded ratification, and labels alone are not ratification evidence, so it is queued here instead of dispatched.

Also needed: the triage brief left the woodcut-ink pack decision open. Option (a) is the one-function fix. Option (b) redesigns how bands are learned, which triage said would raise the item to C4. Please say which option the worker should implement when you ratify.

Intended dispatch: one /work-items:work worker implements the ratified option and opens a draft PR.