acceptance scenario 39 discards the server's output when healthz times out on CI

[jayminwest]

The defect

When acceptance scenario 39 times out waiting for the public-mode warren instance to become healthy on CI, the child process's stdout and stderr have already been discarded, so the failure leaves no evidence of why the server did not come up.

scripts/acceptance/lib/inproc.ts:169-177 spawns the server with stdout and stderr set to "ignore" unless WARREN_ACCEPTANCE_WARREN_STDOUT=1 or WARREN_ACCEPTANCE_WARREN_STDERR=1 is exported. .github/workflows/acceptance-public.yml:42-46 sets neither. waitForHealthz (scripts/acceptance/lib/poll.ts:88-108) throws healthz did not reach a terminal state within 30000ms (last state: ...) and that one line is all the CI log keeps.

The same failure hit three times in the last eight days on main (runs 33132730151, 33561435518, 33799958392), so the flake is real and the 30-second window with backoff introduced in PR #909 did not remove it. Without the child's output nobody can tell whether the process crashed, exited early, or bound a different port.

Fix direction

Capture instead of discarding. scripts/acceptance/lib/self-host-docker.ts:38-43 already spawns with stdout: "pipe" and stderr: "pipe" and reads them with new Response(proc.stdout).text() on exit. Copy that into inproc.ts: pipe both streams into buffers, and when waitForHealthz throws, append the buffered output (the last few hundred lines is enough) to the AcceptanceError message before rethrowing. Keep the existing WARREN_ACCEPTANCE_WARREN_STDOUT and WARREN_ACCEPTANCE_WARREN_STDERR env vars working as the live-streaming option.

Scope

Out of scope: the root cause of the boot failure on CI runners. That needs the diagnostics this issue adds. Also out of scope: changing the timeout again.

Tests: no test pins inproc.ts. Add a small unit test for the capture-and-attach path using a fake child that writes to stderr and exits non-zero. bun run acceptance:public runs the scenario locally against a real boot.

Getting started

AGENTS.md covers setup and conventions, and the "Acceptance harness" section describes scenario 39. Run bun run check:all before pushing. Warnings count as failures.

Tracked internally as warren-f074.

Revised 2026-09-04. The original report asked to raise the 10-second healthz window and add backoff. PR #909 (fe428898) did both: the window is 30 seconds and waitForHealthz backs off from 100ms to 1s. That bullet is retired. The surviving problem is the recurrence with no diagnostics, as described above.

[jayminwest]

Maintainer note: I am away until Monday, September 14, and will not review or merge PRs before then. PRs for this issue are welcome this week. Comment "taking this" here before you start, and put Closes #1237 in the PR body. @luantaraschi has commit access and can merge a linked PR that passes CI when they have time, but please do not expect a turnaround from them. If this is your first PR to the repo, GitHub holds the CI run until a maintainer approves it, which can take a few days this week. Put questions on this issue, not on the PR. I will work through the queue in the order that the PRs were opened when I am back.