Merge queue test trim

Every change in the merge queue runs some of the same tests twice. This change runs them once.

2026-10-02 · PR #15231 (Fede) · on hold until someone reads it

In one line: the queue now runs the whole test suite on one big machine, so the two smaller test jobs it also runs are repeats. Drop the repeats.

Today: each change in the queue runs

Full test suiteone big machine · ~11 min · every test

Tests for what changed~5 min · already inside the full suite

Guard tests6 machines · ~3 min · already inside the full suite

Type check, lint, build~14 min · not a test, still needed

Proposed: each change in the queue runs

Full test suiteone big machine · every test

Tests for what changedremoved

Guard testsremoved

Type check, lint, buildunchanged

What changes

Same safety: every test still runs in the queue, once instead of twice.
Fewer machines: about 7 fewer machines per change in the queue. That's roughly 1 in 3 queue test jobs (Fede's estimate).
Same wait: the queue is still paced by type check and build (~14 min), so changes don't get slower.
Pull requests unchanged: they still run their own tests before the queue.

The one thing to check before lifting the hold

The big-machine full suite has to be reliable, because it becomes the only place those tests run in the queue. Fede reports it has been clean all day. Last night it was briefly reverted, which is why the hold went on.

Details

Today's queue run on a passing change: full suite on one machine 11 min; tests-for-what-changed 5 min; guard tests 6 shards about 3 min each; type check, lint and build 14 min. Proposed: drop the second and third. The hold was placed on 2026-10-01 night because the full-suite-in-queue change had just been reverted at the time; it is now back and running in every queue run.