Require the Node that does not crash mid-suite on Windows - #247

The browser suite's dev server died in about 30% of pinned runs on Windows, taking the rest of the stack with it. "Node crashes in its own socket code" (issue 207) established it was a 0xC0000409 stack-cookie failure inside node.exe, on a socket path, and left one lead untested. This runs it.

What the measurement says

nodejs/node#62991 is this bug, which issue 207 ruled out on the evidence it had. It was filed on exit code 0xDEAD, but a second reporter there records "0xDEAD / 0xC0000409" on v24.15.0+, and it closes on "a non-deterministic Windows-only stack buffer overrun in the TCP machinery that was fixed in [v24.16.0]" — which is what a /GS fail-fast with ws2_32 on every crashing stack is.

The change

engines asks for ^24.16, and .npmrc sets engine-strict=true.

Both halves are needed: engines on its own is advisory, and pnpm installs anyway after printing WARN Unsupported engine (measured, exit 0). With engine-strict, v24.15.0 gets ERR_PNPM_UNSUPPORTED_ENGINE and exit 1. The cost, said in the .npmrc comment: it gates every dependency's engines too, so a dependency whose range excludes our Node now fails the install rather than warning.

Bounded to the 24 line, not left open at >=24.16, because Node stopped distributing corepack in v25 (nodejs/node#57617, SEMVER-MAJOR) — and corepack enable pnpm is the first setup step in the readme, and how packageManager supplies the pnpm version. An open range would have declared support for a Node where the documented setup cannot run.

No .nvmrc: .github/actions/setup/action.yml already reads node-version-file: package.json, so engines is the single source. The test explorer's own engines is deleted for the same reason — it is private, only ever run from the workspace root, and said >=22, which includes the crashing version.

How to check it

There is no product surface to drive, so the walk is the install gate itself, on this branch:

pnpm install --frozen-lockfile

On Node 24.16+ it installs (exit 0). On v24.15.0 it refuses, naming Expected version: ^24.16 and Got: v24.15.0 (exit 1) — that refusal is the whole change working. CI resolves node-version-file: package.json to v24.20.0, the version the ten runs above were measured on.

Also run: pnpm typecheck, pnpm test (1752 tests), pnpm test:scripts (236), and the F3 suite eleven times on v24.20.0.

One thing that misreads badly in a run log, worth knowing before diagnosing the next flake: Playwright kills its webServer at the end of a green run, so supervise prints a server exited (1); stopping the rest and pnpm relabels the child Command "vite" not found — the crash's two lines exactly. Position separates them: on a green run they sit within three lines of 105 passed.

Left as chips

• apps/web/tests/unit/pages/CapturePage.test.tsx failed once under full-suite load and passed alone and on a re-run — a timing flake, unrelated to this change.
• Nothing checks that the four written copies of the floor (engines, readme twice, CLAUDE.md) stay in step. This change found the previous value's five copies by grep; a check in scripts/lib/ would be its own piece of work.

Closes #207

🤖 Generated with Claude Code