Close activity attempts without locking the attempt after task, execution and run #1000

Problem

Lease renewal (DefaultActivityTaskBridge::observeAttempt) locks rows in the order attempt → execution → run → task.

ActivityOutcomeRecorder::record() takes the opposite order on its ordinary path: task, then execution, then run. On the stale, cancelled and terminated branches it then calls closeAttempt(), which runs ActivityAttempt::query()->lockForUpdate()->find(...). If a renewal of the same attempt is in flight, the renewal holds the attempt and waits for the execution, while the outcome holds the execution and waits for the attempt. PostgreSQL and MySQL resolve that by aborting one transaction with a deadlock error.

How reachable it is: since dec2aff, refusedExpiredAttempt() refuses most expired or settled callbacks under the run lock before taking any task or execution lock. So the ordinary stale branch is mostly reached in races where the attempt is still Running with an unexpired lease when the fence reads it. The code path is still live, and the cancelled/terminated branch closes attempts the same way.

Found during Security's review of #901, which removed the same reversal from the new audit append (ActivityOutcomeRejection::record() now reads the attempt without a lock). This issue covers the pre-existing closeAttempt() lock, which #901 does not change.

Acceptance criteria

• No outcome path requests the attempt row lock after holding the task, execution or run lock. Either lock the attempt first, in renewal order, or close it with a guarded conditional update that needs no prior row lock.
• A deterministic PostgreSQL and MySQL interleaving (same process harness as ActivityOutcomeRetentionTest::testLeaseRenewalAndLateOutcomeAuditCannotDeadlock in Record bounded diagnostic history for rejected late activity outcomes #901) covers renewal against the stale and cancelled outcome branches. It fails with a deadlock on the current code.