#110: runSuite shard option — census-index striping + suite-order index on emit

[lann]

What

Closes #110. shard?: { index: number; count: number } on RunSuiteOptions, mirroring runCases' established striping semantics, so a consumer worker pool can cut the serial case loop's CI tail (polymorph-webcrypto's 267s leg → the issue's ~90–120s at 4 workers).

• Striping: case membership is decided on the census index i (the full enumerated list), i % count === index; a case outside the stripe is neither executed nor emitted. Membership is decided BEFORE only/tag filtering — the interpretation that keeps the invariant union of every shard's rows, in suite order, == the unsharded run's rows (pinned by the partition-identity test).
• Suite-order index on emit: the callback gains an optional second argument caseIndex (the same i) on per-case rows — undefined for envelope/terminator — so the parent merges stripes back into suite order. The L4 row schema is untouched (goldens byte-identical); single-argument emit callers are unaffected.
• Sharded envelope/terminator contract (documented on the option): each sharded call still emits its own envelope + terminator — runSuite cannot see sibling shards; the parent discards all but one envelope, merges rows by caseIndex, and writes the single terminator (the issue's stated topology). Returned RunCounts are per-shard; the parent sums.
• Validation: loud, before any instantiate — integers, count >= 1, 0 <= index < count.
• Per-case timeout semantics unchanged; shard absent is byte-identical to today.

Worker pool stays consumer-side, as the issue argues (imports are live functions and cannot cross postMessage).

Tests

ct-runner/tests/shard_test.ts (5): partition identity (union == unsharded, disjoint), suite-order restoration from merged stripes, param validation, shard × only-filter interaction, shard-absent byte-identity.

Gates

just test-ct-runner — 22 passed / 0 failed (goldens unchanged). Full runtime/conformance lanes ran green on the same tree under PR #112's gate pass.

Implemented by a coder subagent against the issue contract; reviewed by the orchestrator (one review addition: the per-shard RunCounts sentence). One flag from the track, for the record: runCases/harness.mjs live in the external polymorph-test repo, so the mirrored semantics come from the issue text rather than a locally cross-checked source.