Make benchmark timeout diagnostics passive before v0.13 retry

[deverman]

Outcome

Make benchmark timeout diagnostics passive before the v0.13 retry so observing a timeout cannot add Bridge requests during recovery.

Scope

• Capture queue and process state immediately when a benchmark call fails.
• Remove diagnostic Bridge health requests and the undocumented recovery-delay environment behavior.
• Preserve the effective passive wait: 13,000 ms by default and for smoke/stress; 15,000 ms for release.
• Do not change production transport, BridgeClient, plugin, MCP/CLI product behavior, query behavior, or the 45-second deadline.

Validation impact

performance — benchmark-only observer behavior; production fingerprint must remain unchanged.

Acceptance journey

A release benchmark timeout immediately records the untouched IPC queue/process state, waits 15 seconds without sending diagnostic Bridge requests, and leaves enough metadata to distinguish a late response from a stranded request.

Validation

• Contract test proves the immediate queue/process snapshot is present and bridgeHealth is absent.
• Full Swift test suite.
• swift run focusrelay-dev validate --impact performance.
• Confirm unchanged production fingerprint, then run live health/semantic gates and smoke before any single release retry.

[deverman]

Implementation completed locally on perf/222-passive-timeout-diagnostics.

Validation impact: performance.

Evidence so far:

• Full swift test: passed; live tests reported skipped through Swift Testing traits.
• New timeout artifact contract: passed; queue/process state is present and bridgeHealth is absent.
• bash -n scripts/benchmark-suite.sh and git diff --check: passed.
• Plugin tree and non-benchmark shipped sources are unchanged from origin/master; changed Swift benchmark files are entirely #if DEBUG. The repository source fingerprint changes conservatively because it hashes all of Sources, including debug-only benchmark files.
• swift run focusrelay-dev validate --impact performance: tests passed, semantic gate failed at project_counts_active_contract_bridge after two 45-second attempts. Final request A0BB9F5B-6204-4060-809E-3B3784FD50E0 was classified stranded_not_picked_up with request present, response absent, lock absent, and redispatch attempted.

This was not a clean live session: process inspection showed multiple FocusRelay servers owned by other active Codex/opencode sessions plus unrelated CPU-heavy Swift builds. Per the issue acceptance journey, smoke and release were not started. They remain gated on a clean session; no timeout or production transport change has been made.