fix(server): keep a stdin error handler after cancelling a run

[TakalaWang]

Why

ServerRunner removes the child stdin error listener in its cleanup for three paths: the runtime preparation stage, native runs and native interactive sessions.

A run can be cancelled, or hit its wall deadline, while the parent is still writing a large request to the child's stdin. The ~15 MB CPython runtime request is one example. When that happens:

1. The child is killed with SIGKILL.
2. The pending write fails with EPIPE.
3. No listener is left to receive the error, so it is emitted unhandled and the whole host process crashes with Uncaught Error: write EPIPE.

A downstream judge (NOJV) hit this under load, when a Python checker run was cancelled at a request deadline. It currently carries a pnpm patch for it.

Change

• In all three cleanup paths, swap the stdin error listener for a no-op instead of removing it: child.stdin.off("error", onStdinError).on("error", () => undefined).
  • This matches the existing handling of late child error events.
  • Node's child close event does not wait for stdin, so this also covers late errors after a normal settle.
• ServerCompiler and the Go stage already keep their stdin error listeners for the child's whole lifetime, so they need no change.
• CHANGELOG.md: entry under Unreleased. This may conflict trivially with feat(runtime): allow Python interactors in interact #93, which adds the same heading.

Verification

• New test in server-runner-preparation.test.ts, one case per site (preparation, run, interaction): cancel, then emit EPIPE on the child's stdin. All 3 fail on main with 'Error: write EPIPE' was thrown and pass with the fix.
• Real-process repro (not committed): a real child that never reads stdin, a 32 MiB run request, and a cancel right after spawn. On main the host gets an uncaught Error: write EPIPE; with the fix it passes on 3 of 3 runs.
• pnpm run typecheck, pnpm run lint, pnpm test (959 passed, 9 skipped) and pnpm run build all pass.

🤖 Generated with Claude Code