Round 7: zero-text frames, escaped/curly titles, plan-driven banner crops

The consultant's review of the shipped set found two defects, both ours, and both fixed
at the source rather than in the output:

1. The banner/thumb crop re-framing detected only the glowing core of the object. Column
   brightness measures 2-6 levels above the background on dark glass, so it boxed 8 % of
   the frame for an object group spanning 35 %+. The detector is now a background-plane
   deviation plus the largest connected components (`plan_banner_crops.py`, detector in
   `probe_band_crop.py`), 0 % false-positive on the frame's own negative space, stable
   across thresholds. The crop plan is frozen as `banner-plan.json` and consumed by the
   renderer; a thin subject is centred at 53.5 % of the banner width with the shortfall
   filled by the art's own edge pixels, one row at a time (a single averaged fill left a
   visible tonal step, tried and rejected). Measured on the delivered 72: 0 clipped,
   silhouette median 47.0 %, centre median 53.5 % (`verify_banner_crops.py`).

2. `art-meta.json` pinned two slugs to the hand-accepted sample frames — the very frames
   that carried the HUD readouts — which also kept them out of the crop plan. Pins removed.

Prompt side: 9 frames still printed glyphs at the consultant's bar (>=12 px and >=2
characters). Every offending noun is measured and now asserted against: a literal `marked
at 28` printed "28"; `storefront kiosk` printed "PRINATA"; `voucher card` printed
"Voucher"; `market booths` printed "10MARKET"; `glass button` printed "OC"; `gateway pylon`
printed "GATELY". 50 forbidden words, 0 hits across 72 captions, and all 9 frames
re-rendered clean.

Also: YAML escapes are honoured in the frontmatter reader and straight quotes become
curly ones (the ZH cards shipped a literal backslash), the evidence line is capped at 65
characters, and the light theme gives dark artwork a hairline border plus a soft shadow
(consultant option A) so it reads as a dark display case instead of a black slab.

OCR gate at the consultant's bar: 72/72 frames clean (16 carry sub-threshold single
glyphs, exempt as geometry noise).

docs/project-state.md

@@ -123,8 +123,63 @@ Everything below is measured, not planned. A fresh session needs no other contex
 7. **Delivery shipped 2026-10-19** (`327d516` → Gitea `main`, deploy green): 72 EN + 72 ZH cards, 144 banner files, 144 thumbs rendered into `public/`; the 144 v1 PNGs deleted; 288 frontmatter fields migrated; the two components fixed. `npm run build` → **251 pages**, 216 distinct image references, **0 missing, 0 stale `.png`**; live assertions afterwards: EN card 200 `image/jpeg`, ZH card 200 (did not exist before), banner WebP 200, thumbs 200 from both listings, v1 `.png` 404, and the ZH listing's thumbs resolve from the base slug (the old helper would have 404'd every one of them). The repo also carries the art inputs — `scripts/og-gen/metaphors.json`, `captions/<slug>.json` (72), a refreshed `prompts_catalog.json` with 72 distinct entries, via `scripts/og-gen/sync_prompts_catalog.py --check`.
 8. **Art weight: 13.4 MB → 15.4 MB (+2.0 MB), and it buys 72 ZH cards that did not exist before.** Per file: EN card 67 KB JPEG (v1 was 71 KB PNG), banner 40 KB JPEG + 30 KB WebP, thumb 7 KB WebP.
 
-**Two observations for the owner** (not defects, both open):
-- The hero banner carries no text overlay, so the art's left 60 % negative space reads as dark band on a dark page; on `your-disk-full-alert-is-lying` the subject occupies ~29 % of the banner width. A banner-side crop toward the subject (the card already scales the art 116 %) would roughly double its presence. Not applied — it changes the framing the consultant accepted in round 4.
+**Round-6 — the consultant's review of the shipped set (2026-10-19).** Two defects were
+real and both were ours, not the model's:
+
+1. **ZH cards shipped a literal `\` escape** — `你的\"磁盘将满\"告警在说谎`. The renderer's
+   frontmatter parser stripped only the outer quotes and left YAML's `\"` in the string.
+   Fixed by unescaping YAML escapes and converting straight quotes to “ ” pairwise.
+2. **The artwork carried HUD readouts after all** — the consultant read `60%`, `100%` and
+   `00%` off `your-disk-full-alert-is-lying`, plus a `10` and a ruler line on
+   `that-dying-ssd-…`. Root cause: the shared style block inherited from the accepted
+   samples (`technical blueprint vector lines`, `status indicator lights`) and the
+   background's `blueprint grid crosshairs and delicate vector coordinate lines`, plus one
+   metaphor that literally said "percentage gauge". **The first delivery's own claim that
+   the art was text-clean was wrong**: the gate flagged `00%`@0.84 at the old bar and it
+   was classified as geometry noise. The gate now runs the consultant's bar (glyph box
+   ≥ 12 px **and** ≥ 2 alphanumeric/percent characters; single glyphs exempt) and
+   `build_captions.py` asserts a `FORBIDDEN` word list so blueprint/gauge/label language
+   cannot return. Three pilot frames re-rendered at the fixed prompt: **0 OCR hits each**
+   (was `SSD` 66 px, `IDENESLOES` 92 px, `00%` 15 px), then all 72 regenerated.
+
+Rulings adopted: banner/thumb **reframing** (option B — subject targeted at 45–50 % of the
+banner width, thumbs cut from the banner; measured 45–47 % for 5 of 8 sampled frames,
+35–44 % for two, ~14 % for one genuinely slender subject — widening the caption `bbox` does
+not enlarge the subject, measured), light-theme **hairline + soft shadow** (option A), and
+the evidence line capped at 65 characters. Cards are unchanged apart from the quote fix.
+Style is unchanged (option: keep the realistic 3D machinery look).
+**Round-6b — the banner crop algorithm, rewritten after review (2026-10-19).** The
+consultant looked at the probe sheet and found the detection box on the glassware frame
+covered only the local highlight. Correct, and the cause was structural, not a threshold:
+
+| attempt | measured |
+|---|---|
+| column mean (what shipped in round 6) | dark glass sits 2–6 levels above background → boxed the glow only: 8 % of the frame for an object group spanning 35 %+ |
+| per-row background sample + global threshold | a left-margin sample cannot describe a horizontal gradient → 39 % on one frame, 69 % on the next |
+| **deviation from an estimated background plane + connected component** (now) | downsample 16×, median-filter the object away, upsample; `|L−bg| > 16 or |∇| > 10`; 3×3 opening kills the grid lines; 31 px closing bridges the object's own dark gaps; union of components within 10 % of the largest. Stable across thresholds 8–20, 0.0 % false-positive on the left margin |
+
+The closing kernel mattered: at 9 px two frames split into components and the largest one
+boxed a sliver (4.2 % and 5.9 % of frame width for ~39 % objects). Zoom is computed on the
+**silhouette** (target 47 % of banner width, centre 53.5 %), the detected box is padded 4 %
+first, and the window is no longer clamped into the scaled frame — clamping is what glued a
+thin subject to the right edge. Shortfall is filled by extending the canvas with the art's
+own edge pixels **one row at a time** (a single averaged fill was tried and left a visible
+tonal step down the right edge).
+
+Detection lives in Python (`C:/AI/ideogram4/plan_banner_crops.py` → `out/banner-plan.json`,
+frozen into the repo by `sync_prompts_catalog.py`, consumed by `render_from_art.mjs`), and
+`verify_banner_crops.py` re-detects the subject **on the finished banners** rather than
+trusting the plan. Measured on the delivered set: **72 banners, 0 clipped, silhouette median
+47.0 %, centre median 53.5 %** (two outliers reported: `hello-world` 64 % — the focal cursor
+is small but its platform spans the frame; `my-backup-was-using-someone-elses-api-quota` 25 %).
+
+Also fixed here: `art-meta.json` no longer pins `docker-reported-exit-0-the-oom-killer-disagreed`
+and `running-production-infrastructure-solo` to `sample_oom_169_00001_.png` /
+`sample_prod_169_00001_.png` — **those two frames were the HUD-carrying ones the consultant
+flagged**, and the pins also kept those slugs out of the crop plan.
+
+**Two observations for the owner** (both now decided by the consultant's round-5 ruling):
+- The hero banner's dead black left half is fixed by the renderer's reframing (above); what remains is that a genuinely slender subject (a thin cylinder) still reads small — that is the base frame's composition, and widening the caption `bbox` was measured not to change it.
 - Several frames are intentionally very dark; at full banner width the subject can read as small.
 
 **Three decisions still open** (asked of the consultant in round 4 §二; the owner may simply pick the defaults):