feat(eval): finalize C2 decisions - exact trajectory selector and max S3 gate

[hsskey]

Why

The variant-only trajectory task and the S3 per-run aggregate were both under-specified, so a mapping or render fault could read as a passing evaluation. This finalizes both eval-design decisions: input ambiguity is removed and output dilution is removed.

What changed

• Trajectory selector (Decision 1): the variant-only task now targets the exact selector size=md,state=default instead of the ambiguous size=md, which matched five state variants of Chip and made each condition build a different subset. The task's variant field narrows the coverage and S3 expected set to one component, and the data-variant coverage rule is now stated in the trajectory instructions as the coverage requirement's consequence.
• S3 aggregation (Decision 2): spec.md 9.5 now states that the gated S3 value is the worst variant's ratio (max), with the mean kept only as a diagnostic, replacing the "provisional pending a maintainer decision" framing. The scoring code already returns max, so no scoring change was needed. max matches the per-variant 0.05 tolerance; the mean would loosen it by the variant count.
• Correctness confirmation (no code change): an empty comparison set already returns NO_COMPARABLE_VARIANT (never a pass), and a variant that did not render is already a coverage miss that fails variantRecall, so overall is false. Tests were added to pin both.

How I checked

pnpm --filter @tokenloom/eval test -- s3 trajectory coverage verdict   # 380 passed
pnpm build && pnpm typecheck && pnpm lint                              # all OK
pnpm verify                                                           # all gates PASS (types..selftest)

• Added test: the variant-only task resolves to exactly ["size=md,state=default"].
• Added assertion: s3Detail.max stays null over an empty set, so a Math.max() default of -Infinity cannot pass the 0.05 gate.

What I left alone

No weakened gate, no vitest -u, no rewritten committed evidence, no extra refactor. This changes the trajectory prompt text, so its promptHash changes and a new comparison set begins; committed run records keep their existing hash. No money spent (fake adapter only), no public CLI contract change.