vitest worker SIGBUS on /dev/shm-constrained hosts with task cache enabled

[TheHolyWaffle]

Describe the bug

On any host with a small /dev/shm (e.g. the default 64 MiB on GitLab Kubernetes runners, or Docker's default 64 MiB), vp run ... test reproducibly SIGBUSes a vitest forked worker as soon as the task's file-access tracking has written enough path records to fill /dev/shm:

Bus error (core dumped)
⎯⎯⎯⎯⎯⎯ Unhandled Errors ⎯⎯⎯⎯⎯⎯
Error: [vitest-pool]: Worker forks emitted error.
Caused by: Error: Worker exited unexpectedly

Probing the child worker env mid-run shows /dev/shm usage climbing fast:

SHM usage: Filesystem      Size  Used Avail Use% Mounted on
shm              64M   60M  4.6M  93% /dev/shm

This matches the 4 GiB shared-memory IPC mapping in crates/fspy/src/ipc.rs (SHM_CAPACITY = 4 * 1024 * 1024 * 1024). Once a child touches pages that can't be backed by /dev/shm, the mapping faults with SIGBUS — and the vitest forked worker process dies mid-test.

Setting cache: false on the task resolves UserCacheConfig::Disabled in vite_task_graph::config, which skips the fspy IPC channel entirely, and the same workload completes cleanly. So this is specifically a fspy::ipc::channel behavior under constrained /dev/shm.

Filed originally at voidzero-dev/vite-plus#1453, closed and moved here since the affected code lives in this repo. Related: #340 (LD_PRELOAD) and the target_env = "musl" gate added in #328 also touch this module.

Reproduction

https://github.com/TheHolyWaffle/vite-plus-sigbus-repro

git clone https://github.com/TheHolyWaffle/vite-plus-sigbus-repro
cd vite-plus-sigbus-repro
npm ci

# Fails — "Bus error (core dumped)" + "Worker exited unexpectedly" after ~10s
docker run --rm --shm-size=64m -v "$PWD":/work -w /work node:24.15.0 \
  bash -c 'npm ci --prefer-offline && npx vp run --filter "*" test --no-cache'

Change the test task in vite.config.ts from { command: 'vp test --run' } to { command: 'vp test --run', cache: false } and re-run — passes cleanly (~49 s). The repo also has a .github/workflows/repro.yml that runs both cases on ubuntu-latest under docker.

System info

• vite-plus: 0.1.19 (also reproduced on 0.1.16). Pre-fspy versions (0.1.14) are not affected on the same workload.
• Node: v24.15.0
• Host for the repro: node:24.15.0 container (glibc, Debian base), aarch64 on Rancher Desktop. Also reproduced on x86_64 GitLab K8s runners with the same node image.
• Docker: --shm-size=64m (matches typical K8s runner defaults).

Failing logs

=== RUN ===
 RUN  /work

⎯⎯⎯⎯⎯⎯ Unhandled Errors ⎯⎯⎯⎯⎯⎯

Vitest caught 1 unhandled error during the test run.

⎯⎯⎯⎯⎯⎯ Unhandled Error ⎯⎯⎯⎯⎯⎯⎯
Error: [vitest-pool]: Worker forks emitted error.
 ❯ EventEmitter.<anonymous> node_modules/@voidzero-dev/vite-plus-test/dist/chunks/cli-api.lDy4N9kC.js:3444:27
 ❯ EventEmitter.emit node:events:509:28
 ❯ ChildProcess.emitUnexpectedExit node_modules/@voidzero-dev/vite-plus-test/dist/chunks/cli-api.lDy4N9kC.js:3011:24
 ❯ ChildProcess.emit node:events:509:28
 ❯ Process.ChildProcess._handle.onexit node:internal/child_process:295:12

Caused by: Error: Worker exited unexpectedly
 ❯ ChildProcess.emitUnexpectedExit node_modules/@voidzero-dev/vite-plus-test/dist/chunks/cli-api.lDy4N9kC.js:3010:35
 ❯ ChildProcess.emit node:events:509:28
 ❯ Process.ChildProcess._handle.onexit node:internal/child_process:295:12

 Test Files  (1)
      Tests  (1)
     Errors  1 error
   Duration  10.13s

Bus error (core dumped)

Passing run with cache: false:

=== RUN (cache: false) ===
$ vp test --run --no-cache ⊘ cache disabled
 RUN  /work

 ✓ packages/heavy-io-test/src/heavy.spec.ts (1 test) 48896ms
 Test Files  1 passed (1)
      Tests  1 passed (1)
   Duration  49.02s

Possible fixes

1. Probe /dev/shm size at startup and either reduce SHM_CAPACITY to fit, or fall back to a file/pipe-backed IPC when the backing is too small.
2. Catch the SIGBUS (or detect the failed write via a sentinel) and degrade gracefully with a warning, instead of letting the test worker die.
3. The CLI --no-cache flag disables cache hit/store but still resolves the task to UserCacheConfig::Enabled, so fspy tracking still runs. Either make --no-cache imply no tracking (what users typically expect in CI) or document the distinction so users know cache: false is the right knob.

Happy to test patches.