[TheGreatAxios]

A webhook's head could miss the quick retry after its mail reached no listener. A reconcile pass that lost the race to save the triage state merged its plan over the stored rows by comparing timestamps, so it could overwrite a head the coalescer had queued a few milliseconds earlier. When the coalescer's send then failed, it found no queued row to mark unrouted, and the head waited for the new-head grace period and the next five-minute pass.

What changed:

• When a pass merges after a conflict, it keeps every row someone wrote after the pass read the state, and it mails none of those heads. Before, it went by whichever timestamp was later.
• The fix is in the merge rather than in failDelivery, so the coalescer stays the only writer of the rows it queues. Marking the head unrouted whatever its status would have left the pass free to overwrite a queued row and relied on a second write to repair it.
• In the chain tests, a timeout in settled now prints the pull request's rows from the triage state.

How it was verified:

• New Postgres test: a reconcile pass plans while a webhook's head is queued, the head's mail then reaches no listener, and the next pass mails it. It fails before the fix, because the head is not sent on the next pass, and passes after it.
• The interleave, reconciler, plan and coalescer suites pass (88 tests). Hub and tests/chain typecheck clean, and oxlint is clean.
• The disconnected chain scenario passed once locally (108 s). It does not hit the timeout path, so the new settled output was not exercised.

Closes CL-11646

[linear-code]

CL-11646

[TheGreatAxios]

When a reconcile pass loses the save race and reloads, it now keeps every row whose updatedAt changed since it read the state, and it mails none of those heads. It no longer compares the two clocks.

• Repro: the new interleave test fails 3/3 against 27729d5 (head 9 is never mailed) and passes 3/3 on 548cb91. With the coalescer's clock 5 ms later than the pass's, it passes on both, so the rule no longer depends on clock order.
• No legitimate pass write is lost: a pass's row loses only to a row someone else wrote after the pass read the state, and only on the save after a conflict. If the pass moved a head to queued or running while the coalescer touched it too, the coalescer's row wins and the pass leaves that head unmailed, because the coalescer is already mailing it. Anything else the pass observed is derived again on the next pass. restoreHead puts back the row as it was before, with its old updatedAt, so a send that was undone still counts as unchanged and the pass's row wins. A head the pass never read (read.get is undefined) always goes to theirs, which is the reported race. Passes on one loop never overlap, so the other writers are the coalescer, the request routes and markBaseMoved, and each of them stamps a fresh updatedAt.
• settled(): artifacts.artifact has tenant_id, title, content and archived_at, and triageStateName(repo) matches the GitHub state key the chain repositories use. A missing artifact parses to [], so the timeout path does not throw on it.
• There are no CI-wait leftovers, and the diff is clean.
• hub and tests/chain typecheck. The interleave, triage-reconciler, reconcile-plan and coalescer suites pass 88/88.

No findings.