feat: stop autoscale waking a pool for a run that will never start

[aicayzer]

Closes #65.

The bug

A GitHub run can enter queued and stay there permanently with zero jobs attached. _rp_autoscale counted queued runs, so it woke the pool, found nothing to do, idled out after RUNPOOL_IDLE_SECS and woke again. Observed in the wild: 29 wakes in a day, spaced 21 minutes apart through the night, against a normal five to eight, for two days.

Nothing reported it, because a pool that wakes and stands down is behaving exactly as designed.

The rule

Known-fruitless runs are subtracted from the queued count, never suppressed as a pool. One dead run in one repository must not blind an org pool to the other fourteen, and every other queued run still wakes it normally.

A strike is earned only when state/pools/<name>.started has changed since that run was last judged, which is the only available evidence that waking for it achieved nothing:

• a tick count would give three strikes in three minutes to a pool that cannot start at all, then refuse it real work once its agents were repaired;
• elapsed time would punish a laptop that slept overnight with work genuinely queued.

A whole-queue fingerprint was the first design and does not survive review: on a busy org the set changes whenever any other run arrives or completes, so a stuck run never accrues consecutive evidence.

Suppression clears itself when the run leaves the queued set, and there is a daily re-arm worth exactly one wake so a run held by something transient, an upstream concurrency group being the realistic case, cannot be held for ever.

Surface

• doctor warns, naming the run, its age and its job count, with the cancel command.
• status says how many runs a pool is ignoring, in both the table and --json, and in --local too since the state is local.
• The notifier fires once per hold at warning, keyed on the run rather than the pool so a receiver deduping per pool cannot swallow the next one.
• RUNPOOL_STUCK_WAKES (default 3, 0 disables), through all four sites of the precedence block.

Also here

_rp_queued_runs discards the response body on failure. gh prints GitHub's error JSON to stdout, so a watch list naming a repository that has been renamed or made private used to hand the caller a line beginning {"message":"Not Found" — which reached an arithmetic expansion and aborted it. Covered by a test case.

The API call count per pool is unchanged; only the page size and the --jq differ.

Verification

tests/stuck-queue-guard.sh is new, offline, and covers 25 cases: the pure rule (including the unchanged-started-stamp regression and the org held-plus-new case) and one autoscale loop end to end with gh stubbed, asserting three wakes then silence, that a held pool still costs exactly one call, and that the notification fires once carrying the run id.

bash -n, shellcheck --severity=warning and all eight tests pass.