perf(ci): guard-tree#m0 runs serial ~23 min in x86-main; shard to <=10 min (PR-flow kaizen) #5427

Evidence: run 38881719243 (PR #5413), x86-main step guard-tree#m0 'Every cargo-free guard runs' = 1404.6s, single largest wall-time step; step 3 of the Sections job. Operator target: 8-16 merged PRs/h; shards must be <=10 min.

Proposal: run the cargo-free guards in N parallel shards (or parallel inside the step) with identical pass/fail semantics and every-failure-reported behaviour (m0 case table must still pass). Workflow change => quorum sign-off by a non-author; no waiver, no third state.

Also measured 2026-10-16: merge_group jobs take ~1 min when reused but queue 26 min for a runner (infra-b3 adding slots).

Refs: aprender-pace PACE-4.