fix(gallery): regenerate the 25 drifted heroes from their code, with contact sheets

[TMHSDigital]

Summary

Closes #200 and does the unblocked part of #202.

• Measure (done in chore(scripts): add measure_hero_drift.py for issue #200 #218): 25 of 79 committed heroes no longer matched what their scripts render.
• Pick one path (Committed hero stills do not match what the scripts render #200 step 2): the PNG-to-webp conversion is now committed, as scripts/render_hero.py.
• Regenerate (Committed hero stills do not match what the scripts render #200 step 3): the 25 drifted heroes and their previews, from the code.
• Contact sheets (Backfill contact sheets (26) and asset sheets (28) for existing showcase pieces #202): built for all 25, plus the 8 matching showcase pieces that had none. The pieces measured and rendered are listed in the table below.

Type

• feat
• fix — committed stills disagreed with the code
• docs / chore / ci / refactor

Changes

• scripts/render_hero.py: renders an entry's README --output command. The script's own framing (exit 10) and asset-quality (exit 11) gates run first. It then writes the hero at 1280×720 and the preview at 1200×675, both webp q90, the quality the showcase scripts already use.
• scripts/contact_sheet.py: writes docs/gallery/contact-sheets/<name>-contact-sheet.webp in the committed format (2610×402; candidate plus the pinned armature-bend, damped-track-aim, bmesh-gear). It prints mean luma, stage (frame-border luma), wedge warmth (R−B over the brightest tenth of the lower half) and saturation per tile. Checked against the committed cart sheet: same size, tiles within encode noise (mean abs diff 0.003).
• 25 heroes and 25 previews regenerated: wave-displace, custom-normals-shade, temp-override-join, treasure-chest, wall-torch, collision-hull-proxy, tavern-stool, gn-modifier-inputs, compositor-glare, bake-normal-high-to-low, damped-track-aim, shipping-crate, park-bench, text-version-stamp, wheelbarrow, campfire, fence-kit, grindstone, anvil, signpost, terrain-scatter, hand-pump, iron-cauldron, light-link-studio, vse-cut-list.
• 33 contact sheets: 17 new, 16 replaced.
• assets/og-card.jpg rebuilt; two of its four tiles were regenerated.
• CLAUDE.md: one sentence naming the three tools under the contact-sheet gate.

Evidence

• live-run-proven: E:\Blender-Developer-Tools\.scratch\blender-5.2.1-windows-x64\blender.exe (reports Blender 5.2.1 LTS).
  • render_hero.py: # 25 regenerated, 0 failed. Every script's own render-path gates passed.
  • Re-measured with measure_hero_drift.py: all 25 now matches at 0.00–0.04% against a q90 re-encode. They were 3.8–77% before.
  • damped-track-aim's re-measure hit Three examples write a relative --output to the drive root, not the working directory #222, a relative --output written to the drive root; that fix is in fix(examples): resolve --output against the working directory in three render paths #223. Its committed hero, rendered through an absolute temp path, equals a fresh render exactly (mean abs 0.00000).
  • python tests/check_gallery_images.py → 158 image(s) ok. build_gallery.py leaves every gallery HTML file unchanged; only the image bytes changed.
• Calibration band re-measured on regenerated files: damped-track-aim renders at 0.1511, not the 0.1722 it had as committed. The band from code is 0.151 … 0.305; armature-bend 0.2787 and bmesh-gear 0.3048 are unchanged.

Contact-sheet verdicts

The criteria are stage darkness, wedge warmth, subject fill, saturation and thumbnail legibility, judged by eye from the sheets, with the measured numbers below. The luma band is 0.151–0.305. Calibration tiles: armature-bend 0.2787 / stage 0.2132 / sat 0.1436; damped-track-aim 0.1511 / 0.1448 / 0.2305; bmesh-gear 0.3048 / 0.2368 / 0.1948.

Sheet	Why	Luma	Stage	Warmth	Sat	Verdict
damped-track-aim	#200 regenerated	0.1511	0.1448	+0.2772	0.2305	calibration member itself: sheet shows the other two
hand-pump	#200 regenerated	0.1666	0.1701	+0.0585	0.0834	pass
street-lantern	#202 (hero already matched)	0.1705	0.1723	+0.0301	0.0983	pass
signpost	#200 regenerated	0.1712	0.1752	+0.1099	0.1187	pass
wooden-bucket	#202 (hero already matched)	0.1729	0.1767	+0.1227	0.0733	pass
wheelbarrow	#200 regenerated	0.1756	0.1645	+0.0796	0.1262	pass
treasure-chest	#200 regenerated	0.1778	0.1871	+0.1744	0.1363	pass; subject reads small at thumbnail scale
iron-cauldron	#200 regenerated	0.1789	0.1715	+0.0220	0.0827	pass
wooden-ladder	#202 (hero already matched)	0.1797	0.1794	+0.1029	0.1197	pass; thin subject, weak thumbnail legibility
hitching-post	#202 (hero already matched)	0.1820	0.1841	+0.1061	0.1188	pass; thin subject, weak thumbnail legibility
anvil	#200 regenerated	0.1826	0.1867	+0.2106	0.1134	pass
shipping-crate	#200 regenerated	0.1833	0.1760	+0.0124	0.1705	pass
park-bench	#200 regenerated	0.1841	0.1697	+0.0496	0.1141	pass
grindstone	#200 regenerated	0.1864	0.1656	+0.1601	0.1120	pass
tavern-stool	#200 regenerated	0.1879	0.1905	+0.0963	0.0821	pass
wooden-barrel	#202 (hero already matched)	0.1884	0.1667	+0.2208	0.1785	pass
fence-kit	#200 regenerated	0.1902	0.1669	+0.3013	0.1343	pass
watchtower	#202 (hero already matched)	0.1924	0.1988	+0.1389	0.1409	pass; subject small in frame
water-trough	#202 (hero already matched)	0.1927	0.1800	+0.0995	0.1303	pass
market-stall	#202 (hero already matched)	0.1983	0.1843	+0.1974	0.1542	pass
terrain-scatter	#200 regenerated	0.2143	0.1924	+0.2201	0.1788	pass
vse-cut-list	#200 regenerated	0.2214	0.1950	+0.0177	0.1025	pass
campfire	#200 regenerated	0.2235	0.1763	+0.0275	0.0455	pass
gn-modifier-inputs	#200 regenerated	0.2281	0.1727	+0.1805	0.2670	pass; near-clipped highlight on the large cube's top face
wave-displace	#200 regenerated	0.2426	0.1932	-0.4296	0.6626	intentional edge-to-edge bleed composition, saturated blue as before
light-link-studio	#200 regenerated	0.2479	0.1919	+0.1567	0.2626	pass
collision-hull-proxy	#200 regenerated	0.2576	0.2042	+0.0021	0.1106	pass
custom-normals-shade	#200 regenerated	0.2750	0.2039	+0.1498	0.2114	pass
text-version-stamp	#200 regenerated	0.3299	0.3254	+0.0799	0.1642	bright backdrop by design; now reads 5.2.1 LTS, the site's primary version
bake-normal-high-to-low	#200 regenerated	0.3359	0.2079	-0.2499	0.1882	bright lilac normal-map card by design; stage in band (0.208)
temp-override-join	#200 regenerated	0.3611	0.2966	+0.4034	0.2167	flag, #224: stage clearly brighter and cooler (border 0.297 vs 0.145–0.237); was already bright as committed (0.338)
wall-torch	#200 regenerated	0.3642	0.2895	+0.4321	0.2622	deviation by design: warm flame pool; luma is information
compositor-glare	#200 regenerated	0.4168	0.1885	-0.0215	0.3211	deviation by design: emissive rings; luma is information

Systemic observation, not a per-piece failure: the showcase props, as a family, are less saturated (0.07–0.18) than the calibration set and often thinner in frame. They read as a consistent muted-wood set rather than any single one being the odd one out. That is a style question for the showcase as a whole, so I'm recording it here rather than failing pieces.

Out of band by design, or flagged:

• By design: compositor-glare, wall-torch, bake-normal-high-to-low and text-version-stamp are emissive or deliberately bright. Every one was already above the band as committed, and each is now darker, closer to the band.
• Flagged: temp-override-join is the one stage-darkness flag, filed as temp-override-join hero: stage renders brighter than the calibration set #224. Relighting its script is a separate content change.

Not done

• Asset sheets (Backfill contact sheets (26) and asset sheets (28) for existing showcase pieces #202 step 3) need a neutral-render mode across the 28 showcase scripts. Backfill contact sheets (26) and asset sheets (28) for existing showcase pieces #202 stays open for that.

Checklist

• Explicit paths only (87 files). The CLAUDE.md edit is deliberate and one sentence.
• Counts and manifest unchanged.
• No new CI check. The tools are authoring-only; the render-path gates they rely on were proven by these runs.
• DCO Signed-off-by: present.
• No credentials or emails. Binary paths are as the template asks.

🤖 Generated with Claude Code