ci: batch macOS test selections

[Yuxin-Qiao]

Summary

• Batch macOS Swift test selections in groups of four while preserving the existing four-way hosted shard split.
• Build test targets once during swift test list, then run each selected group with --skip-build and --no-parallel.
• Keep the historical local retry for non-timeout failures, but make hosted CI fail immediately so a real test failure cannot retry green.
• Isolate timed-out batches one selection at a time and report accurate discovery, selected-selection, first-pass, retry, recovery, timeout, and timing metrics.

Correctness contract

• Selection counts are computed after sharding and --limit-groups.
• Hosted non-timeout failures preserve their first exit code without a retry.
• Singleton timeouts report one timed-out group and exit 124.
• Batched timeouts retry individual selections and report recovery only when every isolated selection passes.
• A local non-timeout retry that then times out falls through to the same isolated-selection path.
• Operator summaries use one-based shard labels consistently.

Observed performance

The contributor's first hosted group-size-four run passed all checks. Its macOS Swift Test steps took approximately 13m22s to 18m17s despite runner queueing. The final candidate also removes repeated SwiftPM build planning after discovery; exact-head hosted timing is tracked by the per-shard summaries added here.

Exact-head proof

Head: 85a0924e7a7851e0dbcfed6ded9f9ee75187f906

• DEVELOPER_DIR=/Applications/Xcode-26.6.app/Contents/Developer make check
• DEVELOPER_DIR=/Applications/Xcode-26.6.app/Contents/Developer make test: 594/594 selections, 50/50 groups, zero first-pass failures, retries, recoveries, or timeouts
• Real strict group-size-four run: 4/4 selected selections passed through --skip-build
• bash Scripts/test_swift_test_sharding.sh: strict hosted failure, local retry, shard completeness/counts, singleton timeout, timeout recovery, retry-to-timeout isolation, skip-build, and discovery-timing cases passed
• Final branch autoreview: clean, no accepted/actionable findings
• Public Model Identifier Gate: PASS; no model-bearing public data in this CI-only diff
• Dependency freshness: N/A; no dependencies changed

Exact-head hosted CI/security is running.

[steipete]

Exact-head maintainer proof for 85a0924e7a7851e0dbcfed6ded9f9ee75187f906:

• Rebased onto current main; contributor's later --skip-build optimization preserved and reviewed.
• make check passed under Xcode 26.6.
• Full local suite passed: 594 selections in all 50 groups, with zero failures, retries, or timeouts.
• A real strict group-size-four invocation passed 4/4 selections using the built test bundle.
• The deterministic harness covers hosted fail-fast behavior, legacy local retry, group-size-four shard completeness/counts, singleton timeout accounting, batch-timeout recovery, retry-then-timeout isolation, discovery timing, and --skip-build wiring.
• Final autoreview is clean.
• Public Model Identifier Gate: PASS. Dependency freshness: N/A.

Hosted CI/security is now validating this exact head; merge remains gated on that result.