[michaelsilva]

Implements the antialiasing direction agreed in #2 without supersampling, extra camera rays, an extra pass, or a larger backing store.

What changes

• scene { shape-rendering: geometricPrecision; }; auto remains the default and emits the previous shader unchanged
• the primary ray remembers its previous distance sample, but promotes it only after the next step proves the ray passed that local minimum without hitting it, or another surface became nearest
• only a passed sample within t * pixelSize becomes silhouette coverage; a ray still converging on an exact hit never becomes an edge candidate
• floor id 0 is never promoted as a shape silhouette
• a passed foreground silhouette can still blend over a farther exact hit
• the grazing object's id is retained, so the saved point is shaded with the correct material
• the whole-scene sphere fast path is widened by one pixel only for geometric precision, so it cannot discard a valid grazing silhouette
• GLSL and WGSL share the generated path
• registry, docs navigation, roadmap and changelog are updated

Decision sequencing

PR #6 is the older contribution and owns decision 139. This draft therefore intentionally does not add a decision entry yet rather than introducing an out-of-order 140 on a branch where 139 is absent.

After #6 lands, this PR should take decision 140, update its roadmap reference, and advance the repository's next-decision pointer in the same change.

Tests in the PR

• compiler contract: default/explicit auto shader identity, syntax validation, @media, passed-near-miss codegen and WGSL lowering
• real-render WebGL2 + WebGPU test with non-flat lighting: every exact sphere hit must remain unchanged while missed silhouette rays gain partial coverage
• real-render WebGL2 + WebGPU test with a red foreground sphere over a larger blue sphere: the passed foreground silhouette must blend over the farther blue hit
• the registry example remains covered by the existing documented-example GPU suites

The precision helper also passes standalone TypeScript syntax checking. An analytic 96x72 sphere trace of the same camera geometry produces silhouette candidates on missed grazing rays and zero candidates on exact-hit rays.

Performance proof before review

Keeping this PR in draft until the requested real-GPU comparison is recorded. The repository benchmark requires visible Chromium on a real GPU.

To isolate the feature cost, compare the committed branch (auto) with the same branch plus a temporary working-tree opt-in on the heavy orrery scene:

node -e 'const fs=require("fs"),p="src/scenes/orrery.gss";let s=fs.readFileSync(p,"utf8");s=s.replace("\nscene {","\nscene {\n  shape-rendering: geometricPrecision;");fs.writeFileSync(p,s)'
npm run bench:compare -- HEAD --scenes orrery --dpr 2
git restore src/scenes/orrery.gss

This measures auto versus geometricPrecision on otherwise identical source. The visual diff is expected; the requested evidence is the before/after frame/GPU timing in bench-results/compare.md.

Vercel currently reports the external-fork authorization gate, not a build result.

Fixes #2

[vercel]

@michaelsilva is attempting to deploy a commit to the Lucasbonomi Team on Vercel.

A member of the Team first needs to authorize it.