CI: conformance-smoke shards hit the 60-min timeout-minutes cap; slowest shard cancels and blocks merges #6456

The conformance-smoke gap-suite shards have timeout-minutes: 60, and the slowest shard is running right up against it — intermittently (sometimes) or consistently (some PRs) exceeding 60 min and getting cancelled. Because conformance-smoke-complete requires every shard to succeed, one cancelled shard fails the whole check, blocking merges on green code.

Evidence (this week)

In every case the completed shards had zero untriaged failures — the code is fine; the shard just doesn't finish in 60 min.

Why it's not concurrency-cancellation

The cancellations happen on runs that were not superseded by a newer push (e.g. #6449's current head with nothing after it), and the durations sit at ~55–60 min (the timeout-minutes: 60 value), so these are job timeouts, not cancel-in-progress.

Cost

A rerun helps only when the shard is borderline (#6444). For a consistently-over shard (#6449) reruns never clear it, so the PR can't merge without an admin bypass. This is now a recurring per-PR tax.

Options

1. Rebalance the 8 shards — the split is by test-name hash, so a few slow tests can pile onto one shard. Balance by measured duration, or split into more shards.
2. Raise timeout-minutes for conformance-smoke (e.g. 60 → 90) as a stopgap.
3. Speed up the slow tests — CI: stop truncating gap-suite output, add per-test summary #796 (closed) noted an O(n²) normalize_output in the harness; if that's still in the diff path it inflates the slow shard.

Not blaming any single PR — this is CI-infra. Filing so it's tracked rather than absorbed as rerun friction. Related: the chronic cargo-test 120-min-cap timeouts.