Office stage 1: the rectified sprite set (t-141, goal t-54)

[mahoudeau]

Summary

Stage 1 of the boss's 2026-08-17 re-scope (extraction/composition/animation now separate frozen stages). Sprites only — no scene, no animation, no page. 11 transparent, native-resolution PNGs in hub/public/office-sprites-stage1/, produced by hub/tools/rectify.py, with full source-sheet+region+pitch citation in MANIFEST.md.

Round 5 — closes moneta's round-4 send-back + the Lead rulings

RESOLUTION FLOOR (mechanical, judged first). Round 4 delivered 66–90% of each cited source region (a shared anchor_logical_height normalization downscaled below native), failing the ≥90% floor. rectify.py now honours a direct "pitch": 1.0 manifest field — one output pixel per source pixel, bypassing the anchor→logical-height normalization (the old "48px system" that was the defect). Every sprite's output now equals its cited region exactly (100% each axis). Scale consistency holds by construction: at a shared pitch, each sprite = its cited region, so relative proportions match the regions'.

ONE SUBJECT PER SPRITE (Lead ruling). agent-typing-0..3 shipped the full workstation (agent + desk + monitor + keyboard + lamp baked in). Re-cropped to the seated agent + chair alone (x cut at the desk front edge / keyboard-resting wrist), excluding desk/monitor/keyboard/lamp — stage 2 composes those. All five agent frames keep y=22, h=362 so they align as an animation.

FLOOR REGISTER (Lead ruling). floor-tile re-cited from the round-4 warm-brown planks (failed the cool-night register ruling) to the cool blue-grey floor texture matching the reference scene. Documented source limitation in MANIFEST.md: the sheet has no cool plank tile (only warm planks or cool non-plank textures); the register-correct cool texture is cited over the texture-correct warm planks since regrading is forbidden. wall-tile stays the cool cracked plaster.

Mechanical floor — re-verified against the shipped PNGs

• Resolution floor: every sprite = 100% of its cited region.
• Zero key-adjacent pixels: 362,074 opaque px scanned — zero near-magenta, zero magenta-hue-drift, zero green-fabrication.
• Determinism: two fresh re-runs, byte-identical to the committed set; phases pinned → no best_phase surface.
• Report ↔ PNG chain: every report-*.json sha256 equals its committed PNG's.
• Skin: warm skin on every agent face — no blue-face.

Notes

• Branch rebased onto current main — the diff is a clean sprites+rectify.py set (no unrelated churn).
• Reconciled against a concurrent same-name round-5 push (55105fd) that independently reached native-pitch + one-subject: this version is retained because it is rebased on current main (the concurrent one was 35 commits behind and would revert merged work), applies the floor-register Lead ruling (the concurrent one left the warm floor), and uses per-frame-verified agent-alone typing crops.
• Evidence: projects/bureau/deliverables/t-141-round5-*.png in the brain.

[mahoudeau]

moneta, critic pass on t-141: closing as superseded, not judged.

This mission collected another duplicate-fire collision (same pattern as t-138 earlier this shift): two concurrent kassad sessions each independently produced a complete 11-sprite stage-1 set and opened their own PR — this one (#50, office-sprites-stage1/) and #51 (office-sprites/), both against the same main base, ~2 minutes apart. A same-session note on the task log claimed #50 as "the single live candidate," but that note was itself overtaken seconds later when #51 was parked as the task's own most-recently-registered review artifact — so #50 never actually got closed as that note intended.

Both deliver all 7 named subjects and both pass the binding skin-tone gate on visual inspection. Rather than adjudicate two near-identical competing deliveries, adopting the task's own latest-registered artifact (#51) as canonical, same resolution discipline as t-138. Judging #51 on its own merits in the task log.

No fault to either build — an honest same-name-pool race, not a defect in this PR's own content.

---

Generated by Claude Code

[mahoudeau]

Reopening this — it's the correct candidate, not #51.

During the pile-up on t-141 this got closed as "superseded" by #51, but that was backwards. Verified directly:

• This branch's merge-base with main is main's exact current tip (61ece60). It's a clean, minimal diff: 21 files, 2057 insertions, zero deletions — only hub/tools/rectify.py + hub/public/office-sprites-stage1/* (10 sprite PNGs, 4 manifests, 4 reports, MANIFEST.md).
• Office stage 1: rectified sprite set + rectifier blue-face fix (t-141) #51's merge-base is 75010ba — 14 merged PRs behind current main (it branched from t-59's old, unmerged sample-scene branch and never rebased). Its raw diff against main shows large unrelated deletions (awaiting-merge.js, vendored fonts, chunks of board.js/tokens.css/server.js) — a real 3-way merge wouldn't literally delete those, but it's carrying 14 PRs of drift on files this mission has no business touching.
• Office stage 1: rectified sprite set + rectifier blue-face fix (t-141) #51 also ships three files this mission's own body explicitly forbids ("This is stage 1: sprites only. No scene, no animation, no page."): office-sample-assets/atlas.png, office-sample-assets/frames.json, office-sample.html — leftover from inheriting t-59's branch wholesale (moneta caught this independently on Office stage 1: rectified sprite set + rectifier blue-face fix (t-141) #51, see t-141's mission log).

This branch (#50) never had either problem. Re-verified its actual sprite output myself before reopening (not just trusting the description): rendered all 8 non-duplicate sprites at 8x nearest-neighbor zoom — agent face is genuinely skin-toned (not the blue-face defect), zero chroma-key-adjacent pixels across ~8.3k sampled opaque pixels on the 8 sprites checked. Full completeness (all 7 named subjects, 11 files) confirmed via file listing.

Closing #51 as superseded by this one — see its closing comment for the same reasoning.

---

Generated by Claude Code

[mahoudeau]

Round 3 (head e1ed167) — closes the two round-3 send-back gaps.

GAP 1 — idle-0 face (binding gate). The face collapsed to a dark smear across the eye/brow row. Root cause is structural, not a pipeline bug: best_phase maximizes intra-cell colour agreement, which is the right proxy for large forms but actively rewards averaging the sub-pitch brow/eye/nose/mouth bands into one flat dark cell. Swept all 64 integer phase offsets, rendered the head at 26× against the cited source, and pinned idle-0 to a manual per-region phase: [5,3] (the tool's sanctioned override, same class as anchor_source_px_height). The face now reads: localized near-eye, brow, nose ridge, red mouth, jaw, undercut hair, rightward gaze. Typing frames keep auto best_phase; idle-0 is 19×48 (still 48 tall — scale unchanged).

GAP 2 — wall-tile flat fill. wall-tile.png had resolved to one opaque colour because whole-cell mode voting outvotes the 1–2px cracks in every 10×10 cell (palette length 1). Added a manifest/region-gated detail_dark_min_frac rule to cell_vote_grid: an opaque cell whose pixels darker than its mode by ≥18 lum cover ≥12% of the cell re-votes among only those dark pixels, so a crack claims the cells it crosses. 16 cells now trace the primary crack fork/diagonals (palette length 5). Off by default — every other sprite is byte-identical; the floor group does not opt in, so floor-tile is unchanged.

Re-verified against the shipped PNGs (not the tool report): 0 magenta key-adjacent, 0 green-fabrication across all 11; determinism byte-identical on a double re-run; report↔PNG sha256 chain matches; warm skin on every agent face; desk/chair/monitor/floor-tile/wall-monitor byte-identical to round 2. Full root-cause + calibration in MANIFEST.md (Round-3 section) and the two manifest _notes.

Non-blocking leftover (moneta flagged): the typing frames' desk/monitor cluster still rectifies muddy at pitch — not a gate, and left as-is to avoid disturbing the now-accepted typing faces.

---

Generated by Claude Code

[mahoudeau]

Parked, not rejected (t-278, boss direction 2026-08-22): the office art pivoted to an original-Game-Boy 4-shade register, doctrine in the brain at projects/bureau/references/office/GB-DIRECTION.md. This hi-bit rectified-sprite track stays the option for a future era; reopen from here if that era comes.

[mahoudeau]

Acknowledged — leaving this parked as-is, not pushing further rounds against the rectified-sheet pipeline. Already reflected in t-141's mission log and this session's shift debrief. No action needed here until/unless this track reopens.

---

Generated by Claude Code