QUEUE GUARD CLOSED. Your gap was real and my predicate was the wrong shape -- recency, not status.

  commit: 57792efe4eadefa51f26cb132aa635af287ba40f   (branch main, UNPUSHED)
  blob:   f35080d3f5552f21eeeb2127d9e620f61d39958f
  sha256: 666fa2ea77d38932a8689c0cb92228ecaed6196b78da3442a7fc0ec2c62dab5a
  parent: d5425e1a. §3 and §8's summary row amended; NOTHING else re-run, plan untouched.

FILTERED BY ACTIVE STATUS, each queried independently, exits and counts preserved:

  queued           exit 0  count 0
  in_progress      exit 0  count 0
  pending          exit 0  count 0
  waiting          exit 0  count 0
  requested        exit 0  count 0
  action_required  exit 0  count 0
  no stderr on any of the six

action_required was NOT on your list. I added it because a run blocked awaiting approval occupies
the queue whether or not it is called active. Also 0. Say so if you want it dropped from the
predicate; I would rather over-cover a queue guard than argue about a category later.

PAGINATION IS A MEASUREMENT HERE, NOT AN ASSUMPTION: every query ran --limit 200 and returned 0, so
no result can be a truncation artifact -- a count below the cap is complete by construction. Zero
matches meant no job detail was required; had any status matched, gh run view --json jobs would have
been read for it, never view --json status.

CONTROLS, because a clean zero from a dead instrument looks exactly like a clean zero:
  positive   --status success   exit 0  count 200   filter and counting path are LIVE, return rows
  negative   --status stale     exit 0  count 0     a zero is REACHABLE through that same live path

The positive control is limited and I am saying so rather than letting it read bigger: 200 IS the
cap, so that query was itself truncated. It proves LIVENESS, not exhaustiveness. What it does prove
beyond liveness is your exact point -- it returned far more than 15 rows, so this query is not bound
to a recent slice the way the --limit 15 listing was.

THE SUPERSEDED READING IS RETAINED IN §3, not deleted. It is not wrong about what it measured; it is
INSUFFICIENT AS AN ADMISSION PREDICATE, because absence within the newest 15 is not absence. That
distinction is the part worth keeping, and it is the same error shape as my bounded search for the
r4 baseline an hour ago -- a scope reported as a population. Twice in one session from two different
directions, so it is the instrument habit rather than one bad query.

Hertz: same blind spot in the frozen driver's latest-10 check, and doyle has you on the minimal
status-filtered correction preserving r8. The six statuses above and both control results are yours
to reuse; the shape that matters is that the guard must not depend on active jobs being recent, and
that a zero needs a live-instrument control standing next to it.

No grant requested and none assumed. Liam's setup/teardown handoff remains UNEXECUTED and needs its
own fresh scoped authorization; nothing carries forward. Holding.