ci: single-retry policy for enumerated contention-flaky files (timeouts only) #1419

Part of #1412 (Track B).

Context

The documented contention-flake pattern (process-spawn tests timing out under ~7x file parallelism, passing in isolation) red-herrings every contributor whose first full run goes red. The policy lives in docs/agents/testing.md prose; CI should encode it.

Deliverables

• An ENUMERATED list of retry-eligible files (start from SUBPROCESS_STUB_TESTS in vitest.config.ts plus the repeat offenders: request-router-open, runner-client, runner-xctestrun, the bench subprocess test) — a shared constant, not a glob.
• CI reruns ONLY those files, once, when their failure is timeout-shaped; assertion failures never retry and fail immediately (the "timeouts only" rule).
• The job summary reports every retried file so retries stay visible, and the list has a ratchet: adding a file requires a comment naming why it spawns/waits.

Acceptance

• An injected assertion failure in a listed file still fails the job on first run.
• An injected timeout in a listed file passes on retry with a visible summary line.
• docs/agents/testing.md updated to point at the constant instead of describing the pattern only in prose.

Review amendments (2026-07-27)

• Every retry-list entry must follow the owned-waiver convention (ADR 0011 style): a trackingIssue URL plus an expiry/review date, not just a prose comment. An expired entry fails the gate until renewed or removed. Retry counts feed the scheduled-lane health telemetry (obs: scheduled-lane health, freshness telemetry, and standard artifact envelope #1430).