[acamarata]

Rows

D-1120 superseding an expired delivery now re-posts the newest status. OutboxClaim marked a seq whose posting claim had expired as superseded without requeueNewest; a poster that died after its HTTP post landed left the host on the stale (possibly green) status. It now requeues the newest posted/posting status, as OutboxPosted and OutboxRelease already do. Test: TestOutboxSupersedingAnExpiredClaimRequeuesTheNewest (fails without the fix: nothing queued).

D-1121 a lost attempt that gets a cancel intent is now settled. Recovery's lost loop got E605 ("cancel intent won") from lost -> retryable, fell into the error branch, and left the attempt (and run) unsettled on every pass. It now ends the attempt cancelled (also on the needs-operator path for non-idempotent jobs), then CloseDeadRuns closes the run. Test: TestRecoveryCancelsALostAttemptWithACancelIntent, both job kinds (fails without the fix with E605: cancel intent won).
Not changed: the row's impact text also names staleChecker treating an unknown coordinator as stale (recovery.go) while docker discovery is daemon-wide, so another home's live container could be rm -f'd. That needs a design decision (container labels carry no home; the process sweep relies on unknown = stale), so it is left for a follow-up.

D-1124 the policy hook now reads the coordinator's store. PolicyRequest gained Store (the coordinator's state.db, set from Coordinator.storePath()); PolicyHook opens it instead of OpenStore(""). Test: TestPolicyHookReadsTheRequestStore (fails without the fix: pin in the home store not read).

Checks (free/ci, -mod=vendor)

• go build ./..., go vet ./...: clean
• go test ./...: all packages pass except two timing flakes under a loaded machine, each passing when run alone with -count=1 and unrelated to the touched files: cmd/agent TestNextStartKillsTheJobOfAKilledAgent, cmd TestRunsWait/the_invoke_deadline_caps_the_wait
• each new test run against the unfixed code: fails as described above