deps: cherry-pick libuv/libuv@e640dc9 - #65118

Neither libuv fix Windows needs is in a released 24.x: the async-close
assertion (nodejs/node#61999) and the 24.16.0 fs-event regression
(nodejs/node#65118) are both on main only. No pin can be green, so float
24.x non-blocking - it goes green on its own once a release carries both.
The 24.19.0 pin this replaces was strictly worse than the 24.15.0 before
it, which predates the fs-event regression and so carries one bug, not two.

Leak detection drops from five jobs to one. Its checks are JS-level and
platform-independent, so the other four repeated a single signal. It gets
--expose-gc via NODE_OPTIONS rather than the test:parallel script, keeping
that a plain test run for developers; without the flag the detector costs
time and measures nothing.

parallel_test.js: the CPU=<n> override is honoured again, and the worker
floor of 4 applies only to the computed default.

Signed-off-by: ulofiai <[email protected]>
PR-URL: #65118
Fixes: #63638
Refs: libuv/libuv#5152
Refs: libuv/libuv@e640dc9
Reviewed-By: Antoine du Hamel <[email protected]>

Signed-off-by: ulofiai <[email protected]>
PR-URL: #65118
Fixes: #63638
Refs: libuv/libuv#5152
Refs: libuv/libuv@e640dc9
Reviewed-By: Antoine du Hamel <[email protected]>

Neither libuv fix Windows needs is in a released 24.x: the async-close
assertion (nodejs/node#61999) and the 24.16.0 fs-event regression
(nodejs/node#65118) are both on main only. No pin can be green, so float
24.x non-blocking - it goes green on its own once a release carries both.
The 24.19.0 pin this replaces was strictly worse than the 24.15.0 before
it, which predates the fs-event regression and so carries one bug, not two.

Leak detection drops from five jobs to one. Its checks are JS-level and
platform-independent, so the other four repeated a single signal. It gets
--expose-gc via NODE_OPTIONS rather than the test:parallel script, keeping
that a plain test run for developers; without the flag the detector costs
time and measures nothing.

parallel_test.js: the CPU=<n> override is honoured again, and the worker
floor of 4 applies only to the computed default.

Signed-off-by: ulofiai <[email protected]>
PR-URL: #65118
Fixes: #63638
Refs: libuv/libuv#5152
Refs: libuv/libuv@e640dc9
Reviewed-By: Antoine du Hamel <[email protected]>

…2824)

## Why

Windows CI has been pinned to Node `24.15.0` since the libuv
file-watcher
assertion crash (libuv/libuv#5010) was discovered. That reason no longer
applies: the fix shipped in libuv#5152, was cherry-picked into Node via
nodejs/node#65118, and first released in **Node 24.21.0** (2026-09-08).

Investigating #2146 (recurring nightly Windows CI
failures), the
dominant signature is a V8 Maglev JIT crash (exit code `3221226505` /
`STATUS_STACK_BUFFER_OVERRUN`, tracked upstream at nodejs/node#62260).
We
already mitigate this for `cedar dev`'s web server via `--no-maglev`
(`packages/cli/src/commands/dev/devHandler.ts`), but:

- the crash still recurs even with `--no-maglev` present — confirmed
upstream on 2026-09-21 that the flag only reduces the crash rate ~5x, it
  doesn't eliminate it
- it also crashes in paths the flag never covers, e.g. `yarn install`
  itself (observed in a Sept 10 RSC smoke-tests run)

One upstream reporter also noted their repro doesn't reproduce on Node
`24.19.0`/`22.23.2` while it does on `24.12.0`, suggesting additional
relevant fixes landed in that range too — all after our `24.15.0` pin.

## What

Drop the Windows-only version override in `set-up-job/action.yml` so
Windows tracks the same floating `24` as every other platform, picking
up
both the libuv fix and any V8/Maglev-adjacent fixes in later 24.x
releases.

This isn't expected to fully eliminate the Maglev crashes (upstream is
explicit that `--no-maglev` alone doesn't), but it removes an
unnecessary
constraint and may reduce the frequency. Worth keeping `--no-maglev` and
extending it to the still-uncovered paths (`yarn install`/corepack,
`cedar serve`, api-server-watch, `cedar storybook`) as a separate
follow-up if failures persist after this lands.

## Test plan

- [ ] Nightly Windows run (`.github/workflows/nightly-windows.yml`)
picks
      up Node 24.21.0+ and completes without regressions
- [ ] Watch #2146 for a change in failure
frequency/signature
      over the following nights

…2824)

## Why

Windows CI has been pinned to Node `24.15.0` since the libuv
file-watcher
assertion crash (libuv/libuv#5010) was discovered. That reason no longer
applies: the fix shipped in libuv#5152, was cherry-picked into Node via
nodejs/node#65118, and first released in **Node 24.21.0** (2026-09-08).

Investigating #2146 (recurring nightly Windows CI
failures), the
dominant signature is a V8 Maglev JIT crash (exit code `3221226505` /
`STATUS_STACK_BUFFER_OVERRUN`, tracked upstream at nodejs/node#62260).
We
already mitigate this for `cedar dev`'s web server via `--no-maglev`
(`packages/cli/src/commands/dev/devHandler.ts`), but:

- the crash still recurs even with `--no-maglev` present — confirmed
upstream on 2026-09-21 that the flag only reduces the crash rate ~5x, it
  doesn't eliminate it
- it also crashes in paths the flag never covers, e.g. `yarn install`
  itself (observed in a Sept 10 RSC smoke-tests run)

One upstream reporter also noted their repro doesn't reproduce on Node
`24.19.0`/`22.23.2` while it does on `24.12.0`, suggesting additional
relevant fixes landed in that range too — all after our `24.15.0` pin.

## What

Drop the Windows-only version override in `set-up-job/action.yml` so
Windows tracks the same floating `24` as every other platform, picking
up
both the libuv fix and any V8/Maglev-adjacent fixes in later 24.x
releases.

This isn't expected to fully eliminate the Maglev crashes (upstream is
explicit that `--no-maglev` alone doesn't), but it removes an
unnecessary
constraint and may reduce the frequency. Worth keeping `--no-maglev` and
extending it to the still-uncovered paths (`yarn install`/corepack,
`cedar serve`, api-server-watch, `cedar storybook`) as a separate
follow-up if failures persist after this lands.

## Test plan

- [ ] Nightly Windows run (`.github/workflows/nightly-windows.yml`)
picks
      up Node 24.21.0+ and completes without regressions
- [ ] Watch #2146 for a change in failure
frequency/signature
      over the following nights

(cherry picked from commit 6923b19)