Every seat voted GO; the panel converged on path-aware routing via crossed-path concessions — and deliberately left one sequencing decision priced on both sides for the operator.
A 15-engineer-hour Friday sprint to restore confidence in a Rust build-graph tool's 12-job GitHub Actions matrix. Forty-one percent of PRs had at least one flaky red and authors waited 18 minutes per run. Five forks were on the table: quarantine flaky suites, deterministically fix the top two, move most testing to a fast PR gate and full merge queue, add process-level nextest retries, or automate triage and reruns. The request required one shippable choice, a concrete watcher-race trace, and a measurable failure invariant.
TempDir and fsync do not prove the watcher observed a rename. The test must wait for sentinel-observed readiness and assert the semantic event, not sleep longer or retry inside the assertion loop.
In-test loops hide which attempt failed and normalize nondeterminism. Nextest retries preserve clean process state and must report PASS_AFTER_RETRY separately.
The 80% of ordinary library changes use the fast gate; watcher, cache, platform, CI, and classifier-uncertain changes run the full matrix pre-merge; every merge-group runs everything.
A 45-minute re-engagement cost was downgraded from arithmetic input to a bounded sensitivity estimate, alongside three more precision challenges in one round. The routing decision survives on stipulated wait times and coverage, not invented precision.
Once 80% of PRs stop running flaky suites pre-merge, aggregate PR-red improves mechanically; '41% became 10%' is invalid across different execution stages and denominators. The routing stratum is observed, so the bias is recoverable by stratification.
Remediation: Report per-test first-attempt rates on merge-group + scheduled stress runs + stratified PR-red; never compare aggregates across execution stages.
Transitive Cargo changes, build.rs, proc macros, generated code, feature flags, renames, synthetic merge commits, and base-branch movement can route a sensitive change to the fast path. False negatives need a dedicated metric and automatic fallback.
A PASS_AFTER_RETRY is always YELLOW and never improves first-attempt metrics; every test must maintain >=99% first-attempt passes over rolling 200; the allowlist expires after 14 days with an owned issue.
The ratified rule was signed verbatim with "no further edits." The open cut-line is recorded as a decision boundary, not chosen — each side carries its own degraded-mode invariant.