ci: keep the crash harness's host records on every run and free its memory first

[stephane-segning]

What this does

The crash harness stays on Android emulator 37.1.11.0 (build 15917651). Moving it to 37.2.12.0 (build 16428233) was tried three ways and none of them worked, so this PR only keeps what those attempts showed was worth keeping:

• host.txt and the kernel log are uploaded on every run (if: always()), not only on a failure. A green run is what a red one is read against.
• A "Free the host" step stops the Gradle and Kotlin compile daemons before the emulator starts. They hold about 4.8 GB on the 16 GB runner (free -m went from 5757 MB used to 1120 MB used on this PR's own API 34 job, in the log).
• The comment on EMULATOR_BUILD records the cause. EMULATOR_BUILD itself is unchanged at "15917651", and the guard that reads Pkg.BuildId from the SDK's own record is unchanged.

Why 37.2.12.0 cannot be adopted yet

The hang is host memory exhaustion, and the emulator process causes it. Its resident size grows with every app launch and never comes back:

API 34, one run	emulator process	lowest memory available	swap
37.1.11.0 control, run 37873735095	3.1 GB at the first sample, 3.9 GB through all 20 repeats, 4.05 GB at the end	9717 MB	never used
37.2.12.0, run 37869721841	3.0 GB growing to 13.6 GB	459 MB	full from the 14th minute
37.2.12.0, Vulkan off, run 37873739944	13.5 GB	468 MB	full from the 14th minute
37.2.12.0, -gpu guest, run 37873744804	13.2 GB	93 MB	full from the 15th minute

On 37.2.12.0 the growth is 50 to 300 MB per app launch, crashed or not: the release-build refusal phase, which never crashes anything, still grows by about 60 MB per launch. On 37.1.11.0 the same phases grow by 0 to 16 MB per launch after the first few. The sampler's smaps_rollup shows the growth is anonymous native heap (Pss_Anon 3.3 GB on 37.1.11.0, 8.0 to 8.4 GB on 37.2.12.0 at the end of API 30), on top of a fixed 2 GB guest RAM mapping. The AVD is already at the emulator's own 2560 MB minimum, so a lower AVD RAM changes nothing.

The "hanging thread" watchdog, the QEMU stalls of 15 to 24 s and the lost runner are what the kernel does once physical memory and the 3 GB of swap are gone. Two 37.2.12.0 runs show it directly:

• Run 37871662706: every harness stage passed (native: 20/20, native: 6/6, PASS release refuses the crash channel), then the runner received a shutdown signal during the final flutter test Gradle build, which is the last memory spike. That run's artifact was lost with the runner.
• Run 37873744804: green, but detected a hanging thread 'QEMU2 main loop' (17.7 s), 'QEMU2 CPU0 thread' (21.4 s) and 'QEMU2 CPU1 thread' (22.8 s) were logged in that same final Gradle build, with 93 MB available.

The three attempts, one change each, all measured on the same sampler:

Attempt	API 34 runs	Result
Stop the build daemons before the emulator (this PR keeps it)	37869721841 green, 37871662706 runner lost	Frees 4.8 GB at the start, but the leak still spends it. One pass at 459 MB available is luck.
-feature -Vulkan	37873739944 green	Same growth, 13.5 GB.
-gpu guest	37873744804 green with three hanging-thread lines	Same growth, 13.2 GB.

API 30 under the same workload (no 20-repeat stage) ends at 3.4 GB on 37.1.11.0 (run 37873735095) and at 8.0 to 8.6 GB on 37.2.12.0 in all three variants (runs 37871662706, 37873739944, 37873744804).

A green run on 37.2.12.0 here means the run finished before the memory ran out, by 93 to 468 MB. That is not a proof the harness works on it, so main stays pinned. Bump EMULATOR_BUILD once the emulator's size stays level in host.txt.

Verification

• actionlint, zizmor --offline (no findings) and yamllint -d relaxed on .github/workflows/crash-harness.yml; Super-linter is green on the PR.
• The squash message parses with @conventional-commits/parser 0.4.1, assembled as ci/commit-message-parse/parse.mjs assembles it, and the commit-message-parses check passes.
• Actions stay SHA-pinned, nothing is interpolated into a run: block, and permissions are unchanged.
• This PR's own crash-harness run, 37875987419, on the pinned emulator: both jobs green. API 34 logs Android emulator version 37.1.11.0 (build_id 15917651), the guard prints Pkg.BuildId=15917651, native: 20/20 runs passed, native: 6/6 runs passed, PASS release refuses the crash channel, 1 test passed, and no hanging thread. The emulator stayed at 4.2 GB at most, 9502 MB of memory were available at the lowest, and swap was never used. host.txt was uploaded by the green run, which is what if: always() is for.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XLy7TiPAxtVjPKEwQ9iwBa

---

Generated by Claude Code