Files

T

kami c19a726455 Stop dropping strip pixels, so panel bboxes tile the page

Answering whether the caption merge breaks the page-space key in
eval/chapter-truth.json. It does not. _merge_plan only groups adjacent
fragments and slice_webtoon cuts at gutter midpoints, so a vstacked panel's
[x, y0, w, sum(heights)] is the exact page region.

The defect is one line up. slice_webtoon skipped any segment shorter than
min_seg instead of absorbing it, so this chapter's 116 panels have a 62px hole
between panel_order 70 and 71 and lose 18px at the strip's end. A merge
spanning that hole reports a bbox 62px short of the page span it covers, and
every consumer mapping a detection back to the page is then wrong by that much,
silently. The 62px held the credits line, found at p_050.png row 338.

A short band is now held and absorbed by the next segment, or by the previous
one when it is last. min_seg still suppresses a tiny panel and no strip pixel
is discarded. The self-check asserts the bboxes tile the strip end to end.

Checks: worker_crop.py self-check passes and fails on the old code, where the
first band starts at 49 instead of 0. ruff check . exits 0.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

2026-08-13 23:41:11 +04:00

4.6 KiB

Raw Permalink Blame History

HANDOFF, 2026-08-13 (ninth session)

Live state is in NEXT.md. This file is only what this session did. The previous handoff is in JOURNAL.md.

Asked

Get up to speed from the eighth handoff, pick the next work, then commit to master. Then: can the merged-panel bug be fixed.

Result

No GPU work. No stage ran. master is at bc19f9e and pushed, plus the crop fix below still to commit.

ARCHITECTURE.md step 0, the measurement spine, is built for the character half and its DoD is met.

Landed

what	where
`eval/chapter-truth.json`, 38 labelled occurrences	new file
purity and fragmentation, plus a `--selftest`	`audit_registry.py`
a short slice is absorbed, not dropped	`worker_crop.py:slice_webtoon`
the design, 4 sections	`decisions/measurement-spine.md`

Step 0, the measurement spine

truth: 38 labelled occurrences, 38 matched an assignment at IoU >= 0.5, 0 unmatched
  LIM SEONHO [b1dd5659]  lead      14/16 = 0.88 = baseline  (wrong: photo_of_lead 1, chibi_of_lead 1)
  character_2b1b12a1     woman_a   13/13 = 1.00 = baseline
  Seonho [f0d4e901]      woman_b     7/9 = 0.78 = baseline  (wrong: woman_a 2)
  fragmentation  lead 1, woman_a 2, woman_b 1              all = baseline

Rows key on page-space geometry. Purity is the largest share of one true person in a cluster. Fragmentation is the count of ids holding one person. So neither needs a character_id. The file survives the re-crop and the /characters/reset that the fifth cycle runs. That was the handoff's open ordering trap.

NEXT.md said 2 of woman B's 9 crops were really woman A and never said which. They are panel_order 31 and 33, identified from p030 and p032.

The crop fix, answering "can we fix the merged panels bug"

There is no merged-panel bug. _merge_plan only groups adjacent fragments and slice_webtoon cuts at gutter midpoints, so a vstacked panel's [x, y0, w, sum(heights)] is the exact page region.

The real defect was one line up. A segment under min_seg=64 was skipped outright, so the 116 panels have a 62px hole between panel_order 70 and 71 and lose 18px at the strip's end. A merge spanning that hole would report a bbox 62px short of the page span. Every page-space mapping across it would then be wrong by that much, silently. The 62px held the credits line at p_050.png row 338. slice_webtoon now absorbs a short band into its neighbour (decisions/measurement-spine.md#slices-tile-the-strip).

Audit fixes, all pre-existing

20 characters counted where 14 are live and 6 are merge losers kept on purpose
the assignment spread keyed on name, so the two characters named Seonho summed into one line of 11
the default worked example was panel_index 7, a panel vision skips. NEXT.md's "panel 7" is panel_order 7, one lower, the office shot with 6 detections
nothing about skipped panels. 41 of 116 are skip=True, four checked and all four correct

Open

Step 0's dialogue half. 30 lines with their true speaker, typed visible | offscreen | narrator | unknown. dialogue is an empty array in the truth file. It needs the chapter watched panel by panel, because no per-line speaker truth exists anywhere to transcribe.
28 of 122 dialogue lines sit on skipped panels, which hold no character at all. 23% of dialogue cannot resolve to a visible speaker under any linker. This is the measured case for offscreen as a fourth speaker_ref kind, and it is not built.
The eighth handoff's five unapplied audit findings are all still unapplied, gemma.py first.

Checks

.venv/bin/python audit_registry.py --selftest   # IoU match, greedy tie-break, purity maths
.venv/bin/python worker_crop.py                 # the slice must tile the strip with no hole
ruff check .                                    # All checks passed

The live audit ran against /data/manga.db in manga-orchestrator and reproduced every baseline number. Both files must be inside the container and a rebuild drops them, recipe in AGENTS.md.

Next command

The fifth GPU cycle, unchanged as the next pipeline work, and now measurable on both sides. The GPU is idle. The workers are down, so it starts with ./start_workers.sh then ./check_stale.sh.

slice_webtoon changed, so the re-crop is now mandatory rather than merely planned, and clear s3://panels/<manga>/<chapter>/panels/ first or crop_webtoon keeps the old images.

Run the audit after the cycle and compare against = baseline. Expect the panel count to drop and one panel to grow by 62px. A row printed as unmatched means the detection moved more than half its area.

4.6 KiB Raw Permalink Blame History