fix(ci): two REQUIRED contexts sat behind a single self-hosted runner, and it went offline

[avrabe]

Merges are blocked repo-wide, and it is not a code failure

Format and Version Pin Sweep — two of the nine required contexts — both
targeted [self-hosted, linux, x64, light], a pool with exactly one runner
(pulseengine-ci-01-8). That runner went offline.

The Format job was assigned to it, sat 600 seconds, and failed having
executed zero steps — empty step list, missing log blob. That is the
signature of a job that never started, not a lint failure.

Branch protection requires the context to pass. A dead pool cannot produce a
pass, so nothing can merge.

Diagnosed against the local oracle first

cargo fmt --check      -> CLEAN
.rs files on branch    -> 0

The branch that "failed Format" touches no Rust at all. So the failure was never
about formatting.

Fleet at the time of the move

pool	online / total
light	0 / 1	← both required contexts here
lean-mem	4 / 4	idle
rust-cpu	7 / 7	idle

Both jobs moved to lean-mem.

The rule this encodes

A required context must not depend on a single-runner pool.

Capacity was never the problem for these two jobs — redundancy was. The
CI-capacity work (#1062) had been optimising the wrong axis for them: it asked
"is this pool idle?" and never asked "how many machines can satisfy this
label?". Recorded at both job sites so a future retarget doesn't put a required
context back behind one machine.

Not changed

The four required contexts on ubuntu-latest, and Clippy / Rivet Validation on rust-cpu (7 runners — genuine redundancy).

Note on the runner itself

pulseengine-ci-01-8 is still offline. This PR removes the dependency on it;
it does not fix the host. If that machine is meant to be in service, it needs
attention independently — and it is currently the only carrier of the light
label, so anything else routed there later inherits the same single point of
failure.

Refs #1062

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[avrabe]

Closing — the premise was fixed at the fleet level, and better than this PR fixed it

When I opened this, light had exactly one runner (pulseengine-ci-01-8) and
it was offline, so two required contexts could not run and nothing could merge.
This PR routed them to lean-mem.

Re-measured today, the fleet has changed:

pulseengine-ci-01-5    online   rust-cpu+light
pulseengine-ci-01-8    online   light
pulseengine-ci-01-10   online   rust-cpu+light

light: 3 online runners can satisfy it   (was 0 / 1)

The dead runner is back and two rust-cpu machines gained the light
label. That is the better fix: it removes the single point of failure for
everything routed to light, present and future, rather than moving two jobs
away from it.

Why close rather than merge anyway

The PR is green (59/59) and lean-mem has 4 runners vs light's 3, so merging
would be marginally defensible on redundancy. But the change carries a comment
that says:

[self-hosted, linux, x64, light], a pool with EXACTLY ONE RUNNER

That is now false. Landing prose that asserts something untrue about the
current fleet is precisely the defect class v0.61 was named for — stale claims
sitting behind a green gate — and it would be landed by the person who spent
that release removing them. Not worth two jobs' worth of marginal redundancy.

The rule survives the PR

The finding was never really "move these two jobs". It was:

A required context must not depend on a single-runner pool. Capacity was
never the problem — redundancy was.

And the sharper half, which is why my CI-capacity measurements missed it: every
sample I took asked "is this pool idle?". light was idle, which reads as
healthy. Nobody asked how many machines could satisfy the label. Idle and
unavailable look identical in a utilisation reading.

That invariant is worth enforcing rather than remembering, so it goes to
RQ-62-CLAIMCHECK (#1062), which already owns required-context routing —
any future retarget should have to show the destination label has more than one
online runner.

No code change needed today. Thanks to whoever relabelled the runners.