[koala73]

Summary

The 12:04Z Resilience-Scores run published batch-wide education source-failure even though the education seed was healthy. Railway had admitted all 196 pre-warmed country scores, and the active Vercel deployment then rebuilt only the ranking aggregate, so no country scorer ran and the pre-fix poisoned cohort was republished.

An authorized ranking refresh now rotates every per-country score before it publishes the aggregate. Concurrent countries share one bounded recovery read after an exhausted seed-metadata read, which contains the staged failure to one country without creating a retry herd. The seeder also builds intervals from the final post-rotation score cohort, so fresh scores cannot ship beside stale intervals. If the forced ranking rebuild fails or exceeds its 60-second deadline, the seeder now stops before interval publication and preserves the previous interval generation.

Validation

• The new regression first failed with the pre-warmed poisoned cohort unchanged and a staged metadata-read failure poisoning a second country.
• The focused resilience suite passes 304/304, including 19/20 healthy countries, exactly one shared recovery read after the exhausted read, direct sequential null-result eviction, and the failed-refresh interval boundary.
• The pre-push gate passes 1,133/1,133 resilience tests, 58/58 seed tests, 107/107 server-handler tests, frontend and API typechecks, edge bundle checks, policy checks, and the remaining scoped gates.
• Sidecar and API handler tests pass 370/370; repository lint passes with existing warnings outside this change.
• The full data suite reached sandbox-only mktemp failures in the pre-push-hook fixture. The exact failing file passes 16/16 outside the filesystem sandbox.

Related: #6510

Post-deploy monitoring and validation

• Owner: WorldMonitor operations.
• Window: the first clean Resilience-Scores cron after deployment, plus the following scheduled run.
• Confirm the deployed Vercel and Railway SHAs include this PR and the runtime manifest remains formula=pc with education enabled. The live seed-bundle-resilience service has recovered from the namespace mismatch tracked in fix(resilience): seed-bundle-resilience crashes every tick — Railway SKIPPED the #6450 build, so the seeder reads score:v25 while production writes v26 #6483, but this PR's interval-ordering change remains dormant until Railway builds the new image.
• Record the forced ranking refresh durationMs on both runs. More than 45 seconds fails the operational gate: investigate the route budget and workload before changing the 60-second seeder timeout, so the cron does not flap.
• Capture the production ranking artifact and verify education is observed for 181 countries; any batch-wide source-failure cohort fails the gate.
• Verify score timestamps precede the matching interval publication and both artifacts use the same formula generation.
• If the cohort fails again, disable RESILIENCE_EDUCATION_ENABLED, revert this change, and redeploy. Do not treat merge, deployment, or an HTTP 200 as production acceptance.

---

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
worldmonitor	Ready	Preview	Aug 12, 2026 2:21pm