feat(skills): add the PolyCSS agent skill package and evaluation

[apresmoi]

Summary

• add @layoutit/polycss-skills, a zero-dependency npx installer that drops the PolyCSS agent skill into .claude/skills or .agents/skills
• own the skill content in packages/skills/skill/ and mirror it to website/public/skill/ with pnpm sync:skill; pnpm check:skill fails on drift in CI
• split the single skill.md into an entry point plus 13 per-topic reference docs (authoring invariants, scenes, shapes, loading, lighting, shadows, textures, controls, animation, performance, three parity, troubleshooting, api index)
• track installs with a content-hash manifest so an upgrade refreshes shipped docs, drops removed ones, and refuses to overwrite files the user edited
• fix the receiver-shadow lift default: it was 0.001 against a documented 0.05, so castShadow + receiveShadow emitted paths that darkened nothing in all three renderers
• add eval/skill/, which runs real agent CLIs in isolated workspaces outside the repo and grades what the scene paints in Chromium, across three tracks: polycss (skill installed), polycss-noskill (the control — same library and contract, skill withheld), and three (an external baseline, not a control)

Results

Withdrawn pending a re-run. The earlier skill-vs-no-skill table was measured
with a harness that gave the intervention an extra "you have the PolyCSS skill
installed — use it" line the control never received, so it compared the skill
plus an activation cue against nothing. Agents were also the outer loop, so
ordering skill-less tracks first only isolated the first agent — later controls
could read an earlier agent's installed skill through a sibling path. Both are
fixed here; the numbers they produced are not trustworthy and have been removed
rather than restated.

The harness itself is verified independently of any paid run: reference
solutions score 46/46 on polycss, 29/29 on polycss-noskill and 29/29 on
three, and pnpm eval:selftest catches 13/13 injected mistakes, so the
graders both pass correct work and fail broken work.

shadow.lift remains a known PolyCSS limitation, unrelated to any score: it is
expressed in world units while the depth conflict it must win is resolved in
device pixels, so below zoom ≈ 1 a cast shadow collapses into its receiver and
disappears. The default camera zoom is 0.65. Documented in shadows.md and
troubleshooting.md with the workaround; a projection-aware fix changes
documented semantics across three renderers and is left for a separate decision.

Scores from this harness are evidence that the skill is sufficient, not proof it
was the only source used — workspaces sit outside the repo but are not
sandboxed, and track order is fixed for isolation, which confounds wall-clock
comparisons. eval/skill/README.md states both limits. Gemini is not covered;
its CLI blocks on an interactive login.

Verification

• pnpm test
• pnpm build:packages
• pnpm build:website
• pnpm test:scripts
• pnpm check:readmes
• pnpm check:skill
• pnpm eval:skill --agent oracle --track all — reference solutions at 46/46 (polycss), 29/29 (polycss-noskill), 29/29 (three)
• pnpm eval:selftest — 14/14 mutations caught