[sds]

Split each supported Ruby/OS combination across three GitHub runners. CI benchmarks improve complete workflow time from 4m 08s to 2m 15s (46% faster); two shards took 2m 51s. Each runner retains the existing local worker pool.

The shard selector deterministically assigns every example to exactly one runner. Sixteen new examples cover partition completeness, balance, ordering, and invalid shard settings. Each runner still verifies its executed example IDs and merges local worker coverage; unique Coveralls flags preserve coverage across all shards.

Add a single final check named CI for repository rulesets. It depends on every test shard and lint, explicitly fails for failed/canceled/skipped dependencies, and finalizes coverage reporting. Move lint into the same workflow so the required check includes it. Require CI instead of individual Ruby/OS checks.

GitHub Actions benchmarks (complete workflow, including setup and reporting):

Shards per Ruby/OS combination	Matrix jobs	Elapsed
1 (baseline)	6	4m 08s
2	12	2m 51s
3 (benchmark)	18	2m 20s
3 (final PR verification)	18	2m 15s

Validation: both benchmark workflows and final PR-head CI pass, including all 18 shards, lint, and final coverage reporting. Logs verify all 10,170 example executions (1,695 on each of six Ruby/OS combinations), with consistent suite identities and complete shard coverage. Local two-shard execution passes all 1,695 examples (the existing 1,679 plus 16 new shard-selector tests). Actionlint and RuboCop pass. The required-check predicate rejects failure/cancellation/skipped results in exhaustive local truth-table checks. Final PR-head verification: successful run on ce3e508. Linux test steps take 9–19s; Windows test steps take 59–96s. No repository ruleset settings were changed; the check is ready to select as CI.