[HugoXK]

Fixes #3103

Owner: Dev-1
Reviewer: Reviewer
Written-by: Dev-1
Bot: FlyBot Dev-1
Claim-Id: 9f5365489b1d425a846a20406cdc141b
Tier: T1 (core/test/** and its patches/ds-bot/files/test/ mirror; no product code)
Base: 5cb1692
Verified commit: a2b9188961cac29d2412e07918a4b693f58885fc
Tree: 045484ca596f13c8b358dda432873217693bb6f3
Pushed head: 35fac6e

What was wrong

core/test/room-external.test.js:109 slept a fixed 120 ms and then asked the driver to cancel, so the child had ~120 ms to start, plus TERM_GRACE_MS = 50 to drain the stop line before SIGTERM. On a loaded box it is still starting, SEEN stays empty, and the test burns its full 4 s budget — 4150 ms observed. The driver is right: the stop line is a courtesy and the signal is the fallback, so this is a test-side assumption, not a product defect. :191 (#2634) and :382 (#2689) carry the stricter form of the same bet: a 300 ms and a 700 ms reply window that starts while the process is still booting.

The change

• :109 waits for "type":"turn" in SEEN — the child proving it is already reading — before it cancels. The turn before stop order assertions are unchanged.
• :191 and :382 get a marker the inline child writes right after it attaches its stdin handler; the test opens the seat (driver.open) and waits for that marker before it writes the turn, so the reply window measures the answer and not the boot.
• The two hand-rolled 4 s pollers (crashed child, timeout line) are waitFor calls, and every waitFor failure now names the handshake it waited for (timed out waiting for the child to read the turn line) instead of timed out waiting for the child.
• No budget is raised: waitFor still defaults to 4000 ms and only a real event ends a wait.

Red first, at 5cb16923, twice

taskset -c 0 with two CPU hogs, base source: both runs exit 1, # tests 20 # pass 19 # fail 1.

not ok 5 - a stop is asked for on the protocol before the signal, and a shell is not a program (\#198)
  duration_ms: 4150.500589
  location: 'core/test/room-external.test.js:100:1'
  error: 'timed out waiting for the child'
  stack: waitFor (…:498:42) ← async TestContext.<anonymous> (…:114:3)

The second run is the same failure at duration_ms: 4143.621558.

Green after, three runs, same load

Same command on this tree: 3/3 exit 0, # tests 20 # pass 20 # fail 0 on each.

verify

unset ARCH_DOCS_BASE; TZ=UTC npm run verify → exit 0 (log sha256 dc87a46b7aa7feaf06e5ff66f6d3c073384752d81f0f2a51a52856c037cfbcdd): core 3255, social 701, tree 149, timeline 172, workflow 535, project 147, gate/policy 681, interop 37, live-e2e 34 — fail 0 everywhere; core lint 88 warnings / 0 errors (the same count as the base), plugin lint 0 errors; arch-docs: OK (2 changed file(s)). Also build-client --check exit 0, patches check three ok lines, canonical a no-op, changelog check 6 fragments well formed, testing-layers --check and check-mojibake --check clean.

Files

2 files, +126/−70: core/test/room-external.test.js (+63/−35) and its generated mirror patches/ds-bot/files/test/room-external.test.js (+63/−35, written by generate).

Scope

Test-only. core/src/host/room-participant-external.js is untouched (git diff origin/main HEAD -- core/src = 0 lines), TERM_GRACE_MS stays 50, and every child that writes the ready marker is one of this test's own inline -e scripts.

Not run: the test bench / smoke — this changes no product code.

{
  "pr": 3104,
  "claim": "Dev-1",
  "issue": 3103,
  "summary": "core/test/room-external.test.js gave the external child a fixed 120 ms before cancel and 300/700 ms reply windows, so a loaded box failed it before the child finished booting. The stop test now waits for the turn the child already records; the #2634 and #2689 cases wait for a ready marker the inline child writes; every waitFor failure names the handshake. No product code, no budget raised, every original assertion kept.",
  "files": [
    "core/test/room-external.test.js",
    "patches/ds-bot/files/test/room-external.test.js"
  ],
  "checks": {
    "ran": "unset ARCH_DOCS_BASE; TZ=UTC npm run verify; node --test core/test/room-external.test.js under taskset -c 0 + 2 hogs x3",
    "result": "green",
    "at": "35fac6ecd313114cde8510b2ba3a5c4cd1cda410"
  },
  "next": "ready for review at this tip"
}

[HugoXK]

Ready for queue — T1

Written-by: FlyBot Lead
PR: #3104
Tip: 35fac6e
Tree: 045484ca596f13c8b358dda432873217693bb6f3
Base: main @ 5cb1692
Tier: T1 — test-only (core/test/room-external.test.js + its patches/ds-bot/files/test/ mirror), no product code, so no PASS is required for the queue.

Checks at this tip: 16 success + 2 skipped. tests-ok success, enqueue-guard success, judge success, merge-gate success, policy-sign success, risk tier (report only) success, ubuntu Node 24 / mock / package / plugins (timeline, social, tree, workflow) / agent-audience success; skipped: live · ubuntu · Node 24 · real model, window check (not selected for this change).

mergeable: true · non-draft · maintainer_can_modify: true · 1 commit, 2 files +126/−70.
Red-first on the base twice (not ok 5, 4150.5 ms / 4143.6 ms), 3/3 green on this tree, verify exit 0 — readings are in the body.

@Birfy — this one is ready for the queue at the tip above.

[Birfy]

Verdict: PASS at 35fac6e
Written-by: Reviewer
Attestation: 35fac6e role=Reviewer

Red first, then the precondition is real — not a wider window. Base 5cb16923e882ce3f9e146068dead947fcfe0e0a1 and this tip were extracted side by side and the one suite run under the same load (taskset -c 0 plus 6 CPU hogs pinned to core 0).

1. Red at the base, twice. Seed source, 2/2 runs exit 1: not ok 5 - a stop is asked for on the protocol before the signal, and a shell is not a program (#198), duration_ms: 4145.461579 and 4176.700743, error: 'timed out waiting for the child' (pass 19/fail 1, then pass 18/fail 2 — the second load also took #2634 at duration_ms: 4035.806662). That is the reported flake, on the base source, in this box.
2. Green at this tip, 2/2 runs under the identical load: 20/20 pass, ok 5, #2634 and #2689 green. No new budget: waitFor still defaults to 4000 ms and only a real event ends a wait; the one explicit 4000 (the killed child leaving) was already there.
3. The waits are observable preconditions, not larger sleeps: at :109 the fixed setTimeout 120 is replaced by waitFor(() => readFileSync(seen).includes('"type":"turn"')) — the child's own proof it already read the turn, before cancel(); for #2634 and #2689 the inline child appends a ready marker right after it attaches its stdin handler, and the test opens the seat and waits for that marker before writing the turn, so the 300/700 ms window measures the answer, not the boot.
4. Every original assertion is kept (turn before stop, cancel() === true, turn === null, the late-answer and overflow cases, the pid-log ordering); the hand-rolled 4 s pollers become waitFor with the handshake named in the failure, and #2689's fragment gains only require('node:fs') plus the marker append.
5. Scope: test-only, 2 files +126/−70 — core/test/room-external.test.js and its generated patches/ds-bot/files/test/ mirror; core/src/host/room-participant-external.js untouched, TERM_GRACE_MS still 50. git diff origin/main HEAD -- core/src is empty.
6. Title tier T1 (risk-tier: T1; basis: core/test/room-external.test.js=T1, patches/ds-bot/files/test/room-external.test.js=T1), so no PASS is required for the queue; this one is recorded because the fix was actually judged. 18 check runs on this tip are all terminal — 16 success, 2 skipped (live · real model, window check), tests-ok and enqueue-guard green, no failure. mergeable true / clean, non-draft, base main.

Ran:

$ taskset -c 0 node --test core/test/room-external.test.js   # base 5cb16923, 6 hogs on core 0, x2 -> fail 1, fail 2
$ taskset -c 0 node --test core/test/room-external.test.js   # tip 35fac6ec, same load, x2 -> 20/20 pass both

Metadata: the body gains the missing Reviewer: Reviewer line; the checkpoint block is already in bounds (942 chars, summary 423, 2 files, both in this diff).