ci: fix a racy resource test, bound the native-image test step, skip linker suites there

[oyvindberg]

Three CI fixes that came out of debugging the failures on #637 and the Windows telemetry. None was caused by that PR, and none is affected by the M11 bootstrap bump in #638.

1. MachineResourcesTest was racy by construction

The compile-deadlock regression fills the governor with eight forks, retunes the budget below what they hold so free memory goes negative, then asserts a zero-memory compile is still admitted. But:

hold <- CountDownLatch[IO](1)   // ← eight forks release it
forks <- (1 to 8).toList.parTraverse(i => m.reserve(...).use(_ => hold.release *> release.await).start)
_ <- hold.await                  // returns after the FIRST fork
_ <- m.retuneMemoryBudget(15000) // races the other seven

So the over-commit the test exists to construct only sometimes existed:

forks reserved	used	free	result
5	12800	+2200	❌ 2200 was not less than 0 ← what CI hit
6	15360	-360	✅
8	20480	-5480	✅ (intended)

15000 - 5*2560 = 2200 is exactly the reported failure. Counting 8 makes the setup deterministic. Every other latch in the file already counts its own fan-out; this was the only one that didn't.

2. The native-image test step bounds itself

timeout-minutes is a GitHub-side reap of the step's process tree, and on windows-latest it does not work. On run 30697246865 the step declared 20 minutes, ran 45m07s, blew the job's own 45-minute ceiling at 55m00s, and the runner was destroyed with both telemetry steps still pending — no bsp-diagnostics-windows-latest artifact at all. That is exactly what #627 added the step cap to prevent.

run-bounded.sh kills the command itself and exits 124, failing the step instead of the job, which keeps the if: always() collection alive.

• Not timeout(1) — GNU coreutils, absent on macOS, two of the five arches here. Done in bash so there is one code path per OS.
• Invoked as bash <path>, matching collect-bsp-diagnostics.sh, since the exec bit doesn't survive a Windows checkout reliably.
• Bound sized from a real run: windows did this step in 13m14s (native-image itself was only 4m57s, so the step is the cost). 18m would have been 1.36x headroom, tighter than the >1.5x this cap was sized for — and runner variance is larger than that, with the same macos native-image job taking 18m14s and 24m28s on two runs. So 20m, backstop 25 — above the 22m our bounds can total, and clear of the job's 45 so the job cap stays the last thing to fire.

3. The platform-linker suites are tagged slow

Nine suites in bleep-bsp-tests drive a real linker (Scala.js, Scala Native, Kotlin/JS, Kotlin/Native): compile a source, then link an artifact — for Scala Native, NIR → LLVM IR → clang → executable at Debug, ReleaseFast and LTO. That time is clang's.

Measured from CI telemetry, and the set is the union across platforms because the figures move a lot:

Suite	Linux	Windows
ScalaNativeAdvancedIntegrationTest	63.5s	139.1s
ScalaNativeAdvancedTestIntegrationTest	75.6s	132.2s
ScalaNativeAdvancedLinkIntegrationTest	69.4s	73.7s
LinkExecutorIntegrationTest	99.5s	18.5s
KotlinNativeAdvancedIntegrationTest	78.6s	1.7s (cancelled)
KotlinNativeIntegrationTest	24.8s	53.6s
ScalaJs / KotlinJs Advanced (×3)	27.6s	40.8s

Tagged slow — the tag bleep-tests already uses for its **IT suites — rather than a new one. It means the same thing and is already wired where it's wanted, so no workflow change was needed for the exclusion: the native-image jobs already pass --exclude-tag slow.

The build job is deliberately untouched. It is the canonical full-suite gate, so every linker is still exercised once per CI run instead of five times, and its 36 **IT suites keep running. Excluding slow there would have cost 17.3 of its 25.4 minutes.

Net: native-image suite time 18.3 → 10.6 min per arch, build unchanged.

Listed by name, not by glob — the cost is a property of what a suite does, not what it's called, and the cheap suites in the same packages mention the same link types. Every entry cross-checked against suite names in two CI telemetry artifacts: no typos, no stale entries.

Verification

• run-bounded.sh: returns 0 on success, propagates a non-zero exit (7 → 7), reports 124 on expiry
• MachineResourcesTest: 19 passed, 0 failed — three consecutive runs
• tag registration: bleep reports Expected slow, confirming both projects feed one tag namespace

Note this does not make the Windows hang not happen — it makes it fail fast and leave the telemetry behind, so the next occurrence can actually be diagnosed.

🤖 Generated with Claude Code