fix(scheduler): idle fill leaves a supersession backoff alone — the owning PR does not close because a slot is free (AGT-4036)

[unohee]

Why

AGT-4036 asked for a deferred claim to stop wearing superseded; that landed in #493 (deferred status). Its DoD — grep 'Task superseded' returns only genuine supersessions — is still violated, from a different direction: the current daemon log has 256 Task superseded lines, 84 of them AX-1481, and the ledger shows AX-1481 at attempt #415, ~25 s apart:

22:36:54 claimed READY→CLAIMED  (attempt 413)
22:36:56 EXECUTING→RETRY_AT {"finalStatus":"superseded"}   ← draft grooming: an open PR owns the planned files
22:37:15 RETRY_AT→READY                                      ← idle fill lifts the backoff
22:37:18 claimed READY→CLAIMED  (attempt 414)

The supersession is genuine and the exponential backoff (retryAtFor, 5 min × 2ⁿ up to 6 h) is set correctly — then filterAlreadyProcessed's durable idle-fill branch (idleLiftable) lifts any future RETRY_AT while a slot is free. A free slot does not close the PR that owns the files, so each lift re-runs the drafter against the same PR: 8 draft + 17 decompose calls, $0.37, for one task that could not start.

What

idleLiftable excludes lastErrorCode === 'superseded' — the same shape as the AGT-4305 infra_error gate, one condition. Everything else (ordinary rejection backoffs, NEEDS_SPEC/NEEDS_ENV, elapsed RETRY_AT) is lifted as before.

Verified

• autonomousRunner.infraError.test.ts: new case seeds a durable RETRY_AT with errorCode: 'superseded' → not lifted, stays RETRY_AT; control case with errorCode: 'rejected' → lifted to READY. Mutant (drop the condition) fails the first. 15 passed (one 30 s hook timeout in an unrelated case at load 90, green on rerun).
• tsc + oxlint clean.

After deploy

grep -c 'Task superseded' stdout.log per hour should fall from ~40 to the number of genuinely blocked issues × their backoff rechecks.