Rerank Takes Nine Seconds

Rerank moved onto the desktop GPU. Same card as the 7B. Protected leases, not a polite queue. One pass took 8.7 seconds. The next, 9.0. That is the cost of a second look on silicon that already has a tenant.

Full questions still completed. Python in 48 seconds. Godot in 45. Not fast. Finished.

A later scoped query sat in the queue with a headroom veto and never acquired a lease. It got cancelled rather than left waiting for a slot that was not coming. Source inspection instead. The earlier times stay. The later one does not get to lurk.