ci: fold obsolete focused gates into timed app-host shards

Closed
TE
25d ago
+44-75

Summary

The broad app-host classifier now fails on:

  • ordinary XCTest assertion failures;
  • Swift Testing failures;
  • missing/zero-test summaries;
  • any app-host crash/restart or timeout marker.

Four shard-5 focused gates still existed solely to work around the old failure-accounting bug from #7471. This PR returns those suites to the normal timing-weighted shard batch:

  • CloudNotificationDismissParityTests
  • AgentRestoreLiveOwnerAdmissionTests
  • GhosttyOptionAsAltModsTests
  • CLISSHSessionAttachAnchorTests

The dedicated workflow steps and focused-exclusion entries are removed together, so every suite still has exactly one owner.

Measurement

Hosted run 35589300329 attempt 1:

  • Cloud notification parity: 15s
  • restore admission: 11s
  • Option/Alt: 10s
  • SSH attach anchor: 10s

Those four dedicated launches cost 46s on shard 5. The run-wide app-host launch audit found 45 xcodebuild test invocations with ~306s total start→first-test overhead (6.3s median / 10.7s p90).

Expected effect:

  • app-host xcodebuild starts: 45 → 41
  • focused full-suite selectors: 60 → 56
  • broad timing-weighted selectors: 2782 → 2786
  • total semantic selector ownership: unchanged (2842)
  • shard-5 focused reservation: 245s → 199s, using the measured 46s removal; the returned suite costs are then accounted by the timing packer.

Isolation kept

This PR deliberately leaves process-global/destructive/crash-prone gates focused: system-proxy state, portal/window registries, numeric locale, renderer-memory measurement, Pi Feed global state, RemoteTmux mirror topology/crash suites, socket/file-store state, and other explicitly isolated workloads.

No retry, parallel-testing, or classifier behavior changes.

Related: #6134, #7471, #12180, #13095.

Validation

  • red-first sharding guard requires these suites to have broad ownership
  • existing strict-suite/workflow equality guard still proves every selector is owned exactly once
  • current classifier tests prove assertion/crash/timeout failures remain red
  • full hosted macOS validation requested

Summary by cubic

Folds four obsolete focused test gates back into the normal timing-weighted shard batch now that the broad app-host classifier catches assertion, Swift Testing, zero-test-summary, and crash/timeout failures.

  • Removes the dedicated shard-5 workflow steps and focused-exclusion entries for CloudNotificationDismissParityTests, AgentRestoreLiveOwnerAdmissionTests, GhosttyOptionAsAltModsTests, and CLISSHSessionAttachAnchorTests.
  • Lowers shard-5 reserved wall time from 245s to 199s; the returned suite costs are accounted by the timing packer.
  • Adds a CI guard that fails if these suites regain dedicated focused ownership or drop out of broad shard discovery; discovery is validated against the repository root.
  • Leaves process-global, destructive, and crash-prone gates focused; no changes to retry, parallel-testing, or classifier behavior.