Check time is predictable: test-times.json measured for every gate file, tsc cached, the gate-list audit before tsc, and pull checks pause while main is red #1797

test-times.json lists 41 of 274 gate files (rest default 5 s), so shards are uneven: slowest shard runs a median 1.57x the average (plan says 584 s, real 99–262 s); tsc runs fresh in all 631 checks (~22 s wall, 45–68 CPU-s in the pod, 836 MB heap vs a 1024 MB cap); ~60 checks failed the gate-list audit after ~25 s of install+tsc; while main was red, 21 heads failed the same 'acting on many asks at once' case in 0.3 h. Done: the full gate refreshes test-times.json for every file; tsc runs incremental keyed by base with heap 2048; the gate-list audit runs first (fails in <1 s); a red main pauses pull checks with one card naming the failing test.

Evidence: audit of 631 check logs on the vxrail over 11.6 h (2026-10-15), 308,335 shard-seconds in total; parsed data in the steward's scratchpad ci/.