feat(eval): wire S3 visual scoring to the evaluation runner

[hsskey]

Why

Before this change the evaluation runner hard-coded s3 = null, so a generated
component was never checked against its reference render. After it, S3 is a real
visual-difference metric with an explicit status, so a mapping or render fault can
no longer read as a passing evaluation, while S1, S2, and variant coverage are
unchanged and gated independently.

Stacked on the variant-coverage work (T1) in #10; it consumes T1's expected-set,
marker, and verdict interfaces. Base this PR on fm/tokenloom-eval-coverage and
merge #10 first.

What changed

• variantReferences maps each expected selector to its snapshot component and
  recorded renderPng by exact variantSelectorOf equality, cross-checking the
  importer's node-id filename rule.
• scoreS3 is pure over PNG buffers and returns measured (worst-variant ratio,
  mean recorded), not-applicable (only when no expected variant has a reference),
  or error (mapping, decode, render, or DOM-disagreement fault). An error
  withholds the gated s3; a variant the output did not render is a coverage miss,
  not an S3 error.
• measureS3 renders only when a reference exists, so a no-reference sample is
  not-applicable with no browser. A RenderPort is injected so tests use
  synthetic shots and CI never launches a browser. A DOM marker list that disagrees
  with the tokenizer fails both S3 and coverage.
• Trajectory rows record s3Status through the same helper (not-applicable for
  the three current samples).
• runs.ts and agent-session-port.ts gain optional s3Status and s3Detail.
• spec.md 9.5 documents the S3 status semantics.

C2 (worst variant versus mean as the gated value) is open. The worst-variant
default is provisional; both values are recorded, so the choice is a one-line
change. No committed run row, report, or reference lock is modified.

How I checked

pnpm build
pnpm lint
pnpm typecheck
pnpm verify --gate scope,types,patterns,determinism
pnpm verify --gate tests
pnpm --filter @tokenloom/eval exec vitest run

All pass: the tests gate reports 1375 tests, 0 skipped, 0 failed (includes the
scoring report recomputation); the eval package runs 353 tests including the new
packages/eval/test/s3.test.ts (22) that cover every T3 acceptance item
(identical/different PNG, no-reference null, each error code, one failing variant
keeps partial ratios, selector matching, recall < 1 stays measured with overall
false, DOM disagreement on both, no browser, trajectory via the same helper).

Full integrated verification with benchmarks and selftest is deferred to the
serialized heavy-validation slot.

What I left alone

No weakened gate, no vitest -u, no rewritten committed run rows, reports, or
reference lock, no change to S1/S2 or coverage semantics, and no product package
outside packages/eval. This does not spend money, make a Figma call, or run a
paid model. The eval internal CLI contract gains optional row fields only.