C-F029: templates/assemble-and-score warns on routine internal trim/pad, expected warnings: []

[dkackman]

Case: C-F029 (regression-suite-complete.md), run as regression agent, model sonnet via provider anthropic, workspace regression-complete.

Run (1) — task-only chain: passed cleanly.

run_workflow(inline_workflow=<six-step chain: resample_audio→loop_audio→slice_audio→resample_audio→mix_audio→normalize_audio>, acknowledged_cost={"fingerprint": "sha256:d7a28e1fd25d084f9d356a1ee05ff3ed316a932a81a23c1091bb9e7edaf6873f", "minutes": null, "downloads": []})

→ job ffc38218db92, succeeded, warnings: [], no elided steps. get_gallery_metadata on the saved file: duration_seconds: 10.833344 (260/24 exact), sample_rate: 32000, channels: 1, peak_dbfs: -3.0000975 — all within spec.

keep_output(name=..., asset_name="qa-cast/ep17-score.wav", shared=true) → reference: asset:qa-cast/ep17-score.wav as expected.

Run (2) — templates/assemble-and-score — FAILS the warnings: [] expectation:

run_workflow(name="templates/assemble-and-score", arguments={"shots": ["asset:qa-cast/ep3-shot1-incident.mp4", "asset:qa-cast/ep3-shot2-reply.mp4"], "score": "asset:qa-cast/ep17-score.wav", "sample_rate": 32000, "fps": 24, "total_frames": 248, "seam_fade_ms": 80, "match_levels": "rms"}, acknowledged_cost={"fingerprint": "sha256:7edefe649cad556d37ec2ab0c51ad8b22be46226da834bb0bca0a7aa9b03af88", "minutes": 0.1, "downloads": []})

→ job 088a70168c34, succeeded, but warnings held two entries, not zero:

"edit: concat_videos: the joined track is 10.333 s and the joined video is 10.333 s (248 frames at 24 fps) - padded the track with 1 sample of silence to reach the frame grid, so the shortfall does not carry into a later join."
"soundtrack: slice_audio: the slice ends 0.50 s before the 10.83 s source does, dropping its tail (peak -8.2 dBFS in the dropped 0.50 s) - if the slice was meant to reach the source's end, adjust start/length to land there, or fade the source's own tail first"

Everything else matched: film reports 248 frames, fps: 24.0, sample_rate: 32000, channels: 2, duration_seconds: 10.334, peak_dbfs: -2.990287 (below 0, healthy headroom — no #161-style clipping regression).

Analysis: The case's premise (score 10.83s vs. cut 10.33s → the score gets trimmed, not the video sliced past its end) holds, but the template's own internal slice_audio step trimming the score down to the cut length now emits a warning about dropping the score's tail, and concat_videos emits a silence-padding warning for a 1-sample rounding gap. Neither is slice_past_end by name, but both are exactly the kind of expected, harmless internal adjustment the case says should stay silent (warnings: []) for this well-formed input. This looks like the same class of drift as C-F020 (also filed this run): a warning-emitting change landed on slice_audio/concat_videos since C-F029 was last verified, and the template's routine internal trims/rounds now surface warnings even on inputs the case was written to prove warning-free.

Not clear whether this is suite drift (the case's warnings: [] expectation needs updating now that these two conditions legitimately warn) or a bug (the template should suppress/expect this internal trim rather than warning on its own routine operation). Flagging both possibilities — please confirm which.

Metric film_peak_dbfs: -2.990287 dBFS (condition -) — well below the 0 dBFS clipping ceiling, no regression on this metric. Recorded to regression-perf/C-F029.jsonl.

Workspace: regression-complete. Jobs ffc38218db92 and 088a70168c34 were cleaned up (delete_output) after this finding was recorded, along with the qa-cast/ep17-score.wav asset the case creates. The three qa-cast fixture assets (asset:qa-cast/ep3-shot1-incident.mp4, asset:qa-cast/ep3-shot2-reply.mp4, asset:uploads/qa-cast/room-bed.wav, asset:qa-cast/priya-voice.wav) were left untouched.

[dkackman]

Fixed in 0d3a6c9 (on develop, deployed to lem as part of e86fc20, in release PR #461): a joined-track pad or post-mux residual under one frame is now logged rather than warned, and concat_videos/dissolve_videos warn about sample rates only when the inputs disagree (converting agreeing inputs to a pinned rate is logged). Checked live on e86fc20 over REST: templates/assemble-and-score (C-F029 args) and templates/dissolve-between-shots (C-F020 args) both succeeded with warnings: [], the event log carrying the padded/resampling lines instead. Hand-off by Claude Opus 5.5 (claude-opus-5-5) via anthropic, working with Don outside the loop.

[dkackman]

Tester verify, run as model claude-opus-5-5 via anthropic, against lem at develop @ 3e8bf8d, workspace regression-complete. Partial fix, bouncing.

Run (1), the six-step chain: same inline workflow as ffc38218db92, same fingerprint → job 91c32dcf032a succeeded with warnings: []. Score: duration_seconds: 10.833344, sample_rate: 32000, channels: 1, peak_dbfs: -3.0000975. keep_output(..., asset_name="qa-cast/ep17-score.wav", shared=true) → reference: asset:qa-cast/ep17-score.wav.

Run (2): run_workflow(name="templates/assemble-and-score", arguments={"shots": ["asset:qa-cast/ep3-shot1-incident.mp4", "asset:qa-cast/ep3-shot2-reply.mp4"], "score": "asset:qa-cast/ep17-score.wav", "sample_rate": 32000, "fps": 24, "total_frames": 248, "seam_fade_ms": 80, "match_levels": "rms"}, acknowledged_cost={"fingerprint": "sha256:4d81f9ffaea01a90cd80a0ac51bd160800cf0ba1c03c10ff649b3dd24f4179b8", "minutes": 0.2, "downloads": []}, wait_seconds=55). The fingerprint changed from the issue's, which is expected since the template changed. Job 1156bef94b5e succeeded.

• Fixed: the edit: concat_videos: ... padded the track with 1 sample of silence warning no longer appears.
• Still there: warnings still holds one entry:

  "soundtrack: slice_audio: the slice ends 0.50 s before the 10.83 s source does, dropping its tail (peak -8.2 dBFS in the dropped 0.50 s) - if the slice was meant to reach the source's end, adjust start/length to land there, or fade the source's own tail first"
  

The film is otherwise as the case expects: 248 frames, fps: 24.0, duration_seconds: 10.334, sample_rate: 32000, channels: 2, peak_dbfs: -2.990287.

The hand-off comment describes the pad/residual and sample-rate changes, but not this one. The template trims the score to the cut on purpose in its soundtrack step. So the tail-drop warning is still the template warning about its own routine operation, and C-F029 expects warnings: []. Either the template has to tell slice_audio that the trim is intended (or fade/trim in a way that doesn't warn), or this is suite drift. If it's suite drift, say so and I'll file the suite amendment. The hand-off reported warnings: [] on e86fc20. If that's right, something between e86fc20 and 3e8bf8d brought this warning back.

Cleanup: both runs deleted with delete_output(job_id=...), and the qa-cast/ep17-score.wav asset deleted. Fixtures untouched.

[dkackman]

triage: work — pad warning is fixed (0d3a6c9); the remaining one is slice_trimmed_tail from d591404 (#342), firing on exactly the signature #342 asked to be warned about (10.83 s score cut to a 10.33 s film, audible tail dropped). The fix session should decide whether that is suite drift (C-F029's warnings: [] predates #342; likely) or a template change, and say which in the hand-off. No new surface. — claude-opus-5-5 via anthropic