test(mcp): report why the server did not answer, instead of hanging on it (#350)

[Ercaner1988]

Fixes #350.

rpc() gave up after a fixed 15s, returned whatever had arrived, and dropped the child's stderr. Callers then dereferenced into undefined, so a slow server and a crashed one produced the same Cannot read properties of undefined (reading 'result').

Chasing that turned up something worse than the bad message, which is why this PR is a little larger than the issue implies.

A crashed server did not fail the run — it hung it

The last two lines of the helper were:

  child.kill();
  await once(child, 'exit').catch(() => {});

That works only while the child is still alive when kill() is called. If the server has already died on its own — a bad import, a missing native build, a throw during startup — exit has fired and will not fire again, so the await never settles. Isolated, without any of graft in the way:

$ node --input-type=module -e "
  import { spawn } from 'node:child_process';
  import { once } from 'node:events';
  const c = spawn(process.execPath, ['-e', 'process.exit(3)']);
  await new Promise(r => setTimeout(r, 300));   // let it die first
  c.kill();
  await once(c, 'exit').catch(() => {});
  console.log('never printed');
"
Warning: Detected unsettled top-level await

End to end, with the server replaced by something that writes to stderr and exits 3:

	result
before	still running when I killed it at 120s — not one of the file's five tests had reported anything
after	the whole file finishes in 3s, the failing test reporting in 0.1s

After

Error: mcp server answered 0 of 3 in 0.1s — it quit first (exit code 3, signal null)
--- its stderr ---
native build missing
    at rpc (test/mcp-server.test.ts:69:11)

and, when the server is up but silent (deadline dropped to 100ms to force it):

Error: mcp server answered 0 of 3 in 0.1s — it was still running
It wrote nothing to stderr.

Those two are now different sentences, which is the whole point of the issue.

What changed, in one helper

• Throws instead of returning short. How many of how many arrived, and how long it waited.
• Keeps the child's stderr and puts it in that error. It was already piped; it was just never read. This is the line that turns "the tool list was empty" into "the native build is missing".
• Waits on a close promise created at spawn time, so it settles whether the child crashed on its own or was killed at the end. This is the hang.
• Stops waiting when the child quits. No point sitting out a deadline nothing can arrive before — and it is what makes the crash case 0.1s rather than a full timeout.
• The deadline goes 15s → 60s, named. Only reachable now by a server that is up and silent, and the loop leaves the moment the replies are in, so a generous number costs nothing in the passing path. It is also the number that was actually flaking: ~2s of work against 15s sounds like ample headroom and is not, once the suite is running ~30 files in parallel.

close rather than exit throughout: it fires once stdio is drained, so a server that answers and then quits is not misread as having quit without answering.

Verified

• test/mcp-server.test.ts: 5 of 5 pass, 1.5–3.7s each.
• Full suite: 1221 tests, 1208 pass, 8 fail, 5 skipped — the same 8 that fail on upstream/main on this machine, which are Savings are under-counted ~1000x outside en-US: the footer is written with toLocaleString() and read back with a comma-only regex #338 (fixed separately in fix: pin rendered numbers to one locale so graft can read its own savings footer (#338) #345) and have nothing to do with this file. (I measured main at 1220/1207/8, but at de8456e, which is older than this branch's base. The one extra test here is the brain build worker …, added by 270edcc in between. This PR adds no tests.)
• Both failure paths above were produced by temporarily editing the helper locally (deadline to 100ms; spawn target to a crashing one-liner); neither edit is in the diff.

Environment

Node v24.16.0, Windows 11 x64, on f9e6539 (0.18.0 plus the brain commits after it).

🤖 Generated with Claude Code

[trailhq-graft]

🌱 graft blast radius

Nothing outside this diff depends on it. 0 areas changed; no indexed dependents at depth 2.

_{graft blast · origin/main...HEAD · depth 2 · 1 changed file}

Open the interactive graph → — click an area to see its dependent symbols at file:line.