fix(runtime): stop waiting for stream EOF when exporting runtime files

[TakalaWang]

Why

Python runtime preparation occasionally stalled until its 300 s deadline when several engines ran at once (about 4 in 100 cold starts downstream in NOJV, with 2 engines per process). The stalled server-runner-stage.mjs child sat idle in Instance.wait(), with no Wasmer worker threads left.

The cause is a race in @wasmer/sdk 0.10.0:

1. Command.run()'s dedicated task sends the exit code.
2. It then calls thread_pool.close() before it drops the WASI runner that owns the stdout/stderr pipe writers.
3. The scheduler handles Close by dropping every WorkerHandle, which terminates its Worker.
4. When that termination wins the race, stdout and stderr never reach EOF, so Instance.wait() never settles.

Instrumented stalls showed the complete 10,695,683-byte archive already delivered on stdout, with both streams still open. CPU contention from concurrent engines widens the race window.

Change

• Add readRuntimeFilesExport():

  • close stdin;
  • read stdout until the self-delimiting WOJFS002 archive is complete;
  • cancel both streams;
  • report stderr if stdout ends before the archive is complete.

  The parent still checks the archive SHA-256.

• Use it in the server runner stage and in the browser runner Worker, which had the same Instance.wait() export path. Free the instance afterwards.

Verification

• New unit test: a complete archive on stdout/stderr streams that never close resolves. A read-to-EOF implementation times out on the same test. An incomplete archive reports stderr.

• Repro harness, before → after:

  Scenario	Before	After
  2 engines per fresh process, Python run	3/50 stalls	0/50
  Python + C++	2/40	0/40
  6 concurrent export stages	49/300	0/300

• pnpm run typecheck, pnpm run lint, pnpm test (958 passed, 9 skipped) and pnpm run build all pass.

• Not covered: the browser path is covered by typecheck and unit tests only, not a real browser run.

• Still open: a rarer C++ compile stall seen once downstream did not reproduce here (0 in 160 compiles). The clang stage doesn't use Instance.wait(), so it may have a different cause.

CHANGELOG.md gets an Unreleased entry, which may conflict trivially with #93–#95.

🤖 Generated with Claude Code

[TakalaWang]

@JacobLinCool Rebased onto main after #93 and #95. CI is green and it merges cleanly. Ready for review. Merging it first also clears the unrelated Python stall that failed #99's latest CI run.