The merge queue never re-forms a batch around the same failing head: suspects are judged alone while the rest lands (one red pull stalled 70+ for ~45 min) #4999

Probe, 2026-10-22 03:27Z: one bad pull blocked the merge queue for about 45 minutes while 70-83 approved pulls waited, and only 2 landed in 20 minutes with 5 of 8 check slots idle.

From the [queue] log, every batch since about 02:45 started with #4812 (#4812 then #4814, #4817, #4825 …). Each went red: 7, then 2, then 4, then 5, then 7 again. After each red, the queue re-formed the batches with the same head, and only at 03:27 gated #4812 alone.

Done when:

• Never the same head twice. After a red batch, the pull common to every red prefix (here the head) is gated alone in parallel with the rest of the batch gated without it, so the good pulls land while the suspect is judged.
• Never re-form a batch around a suspect. A pull that was in two red batches is not put at the head of a third. It waits for its solo verdict, and the tick holds it if it fails alone.
• Use the free slots. When check slots are free and the queue has more than 20 waiting, gate disjoint batches in parallel, up to queue.depth.
• Test: a fake queue with one bad pull at the head and 30 good ones behind. The good ones land within two gate durations, and the bad one is held after its solo red.

Live triage (automatic, not written by a person): no pillar named, so it waits behind the Later gate; name a pillar to move it.