[TheGreatAxios]

When a helper run ends, the hub tells its coordinator. Before this change:

• one unreachable coordinator blocked every other project's notice, because each pass retried the oldest unsent notice first;
• a notice for a project with no active coordinator, or whose coordinator had ended but was still bound, was marked "not notified" and never sent.

Now each ended run carries its own retry state (next_attempt_at, attempts, migration 0036), following Interchange's dispatch-store pattern:

• mail errors of 500 and above retry on Interchange's dispatch delay (capped at 30 s);
• no coordinator, an ended coordinator, or a 409 from the mail route waits with backoff (capped at 5 minutes) and delivers once a coordinator can take it;
• other refusals still settle as not notified;
• the upsert guard keeps mail-once.

Closes CL-11895

Verification

• lint, format:check, typecheck, test (70 pass).
• New e2e test/e2e/specialist-end-notices.test.ts on the throwaway compose DB, asserting stored rows and recipients. 3 tests: a failing coordinator doesn't block another project; an unbound project gets its notice after a bind; a failed but still-bound coordinator's notice goes to the next one. main: 0/3 pass. With this change: 3/3, run 3 times.
• Independent review, including a concurrency check (two reconcilers): each coordinator mailed exactly once. Its findings (the ended-coordinator case, test independence, retry cap) are fixed.

[corbits-triage]

Triage: Unsure

Asking a maintainer about:

• changes sensitive paths: schema: apps/hub/src/db/migrations/0036_ended_run_retry.sql, schema: apps/hub/src/db/migrations/meta/0036_snapshot.json, schema: apps/hub/src/db/migrations/meta/_journal.json

Changed since the last verdict (Fail at bdd4776): 1 new commit.

[corbits-triage]

Triage requests changes:

• Conventions: added code matches a code pattern

Nothing triage asked to change still fails, so this request no longer stands.

[TheGreatAxios]

Verified against a real throwaway Postgres 17: the 3 new e2e tests in test/e2e/specialist-end-notices.test.ts pass, and all 3 fail with specialist-ended.ts reverted to main (schema and migration kept). bun run test: 70 pass.

[TheGreatAxios]

Closing: this rebuilds retry and backoff that Interchange already does natively. Once the hub accepts a mail, Interchange queues it and retries delivery. A coordinator that refuses mail is a fault to surface, not something to retry around. The real bug, one failed notice blocking every other project's, is redone leanly in CL-11918 with no migration and no retry timers.