[cramforce]

Three changes at the Node island boundary, each a place where a program with islands answered differently from Node. They are emitted or linked only in a program that uses an island; the bridge protocol does not change (ABI 20).

What lands

• Clear timers across the Node island boundary #3182: timers are cleared across the boundary. clearTimeout, clearInterval and clearImmediate of an island's timer in compiled code call Node's own function (it did nothing before, so an island interval kept the process alive). Island code that clears a program timer clears it: in a program with islands the program's timer ids start at 2^52, so a number names one timer whichever side made it. A timer an island package declares stays the island's object; storing it in a variable typed as a timer is a compile-time refusal by name. Immediates take their ids from the timer sequence, which also fixes clearTimeout(immediate) cancelling a timeout with the same number in programs without islands.
• Hold a child process a Node island made in a native ChildProcess slot #3348: a child process an island started can be held at the ChildProcess type. The slot holds a child with no state of its own that names the island's object: listeners are registered on the island's emitter, properties and methods are the island object's, the streams are the island's streams. Adoption happens at the checked conversion. Limits are on the limitations page ("Child processes from Node islands").
• The working directory is shared (the remaining half of Share process.exitCode and the working directory with Node islands #3189). Node caches process.cwd() until its own process.chdir runs, so after the program changed directory, island code kept resolving relative paths against the old one (conformance entry packages/process-cwd). In a program with islands process.chdir goes through the island unit, like the exit code: the runtime's change, then Node changes to ".", which drops its cached answer. The exit-code half of Share process.exitCode and the working directory with Node islands #3189 is on main since One registry for process events and one process.exitCode across a Node island #3166.

Changes made on the train

• The open review thread of Hold a child process a Node island made in a native ChildProcess slot #3348 (scr_node_island_child.c, off during an emit) was real and is fixed here (c408bae144). A listener that an earlier listener removes while the event is being emitted was not called in that emit; Node calls every listener registered when the emit started. With Hold a child process a Node island made in a native ChildProcess slot #3348 as it was, tests/corpus/9890-node-island-child-off-during-emit prints a c | a c | a c where Node prints a b c | a c | a c, and a listener removed and registered again during an emit was lost for good (a | a | a against a b | a b | a b); 20 of 20 runs. off now marks the registration as removed and leaves the closure with the function.
• Two merges that did not build on main (27f0d064e0): the timer clear called the island helper by its old name; the adopted child and main each added nodeIslands to the dyn emitter's host.
• A throw decision for nodeIsland.clearTimer and nodeIsland.chdir, which the builtin guard asks for.

Validation

Full round on 90063fc94a (this branch before its last merge of main) against main 6642a17b5f:

• Typecheck passes. Lanes: plain 25,188 passed, 0 failed; sanitized 21,755 passed, 1 failed. The one is the 180 s limit of npm-static.test.ts "commander compiles statically and byte-matches Node" under a full lane; alone on the same tree it passes in 77.7 s, and the same limit was hit in two unrelated validation rounds today.
• Both native bootstraps ran in that round: plain (commands, rebuild) and sanitized (commands, rebuild-frontend, rebuild-emit).
• Zero-cost gate over the whole corpus: 5,569 identical, 0 differ, 347 fail on both sides, 246 runtime-only, 1 build-error.
  • The 246 runtime-only rows are the corpus programs that link the setImmediate code: Clear timers across the Node island boundary #3182 removes the separate immediate sequence from scr_async.c. IR and LLVM are identical on all of them, no function is resized, the size changes by -48 to +24 bytes (50 unchanged, median +8), and no benchmark is among them. This is the exception recorded for Clear timers across the Node island boundary #3182, re-measured.
  • The build-error row is node-island-timers-typescript, a program this train adds: the gate has no libnode pack, so the candidate stops there, as every island program does on both sides. It builds and passes in both lanes.
• Compile time: benchmarks +0.55% [-1.29%, +1.53%], the compiler compiling itself +0.05% [-1.25%, +0.51%], fixtures +0.08% [-0.28%, +0.45%]. Run time: 73 workloads, all neutral. Test262: 0 regressed.
• Island conformance lane in check mode: 438 passed, 0 failed (the list loses packages/process-cwd, nothing else moves). Island value matrix: 239 of 239, and its update mode changes nothing. The weak-keys fixture passes in both lanes.
• The twelve child-process island programs against Node, 50 runs each: 600 of 600 plain, 600 of 600 sanitized; under ThreadSanitizer, 20 runs each: 240 clean. The thirteen worker island programs: 650 of 650 in each lane, 260 clean.

The head adds a merge of main 59950fefd3 with no conflict. On that head: the build, typecheck, guard tests, lint, format and the island tests; the gate's identity and the conformance lane on it are posted as a comment when they finish.

Not in this train

#3252 (silent copies), island-native-twins slice 1, #3043, #3124 and the shared globals branch: each needs its owner's restack on current main first.

Supersedes #3182, #3348 and #3189.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
scriptc	Ready	Preview, v0	Nov 8, 2026 3:48pm UTC

[cramforce]

HOLD: do not merge yet.

The review bot left two findings on this train's child-adoption member (packages/runtime/src/scr_node_island_child.c): at line 260, a close listener dispatched through child.onDyn is released by the cleanup hook before its first invocation, so a rest-parameter close callback never runs; at line 708, pausing or resuming an adopted child's stdout or stderr changes only the native wrapper, leaving the island stream flowing and its reported flow state wrong. Both are being reproduced under Node, on main and on this head. This hold is lifted here when each is fixed or shown not to apply.

🤖 Generated with Claude Code

[cramforce]

Runs on this head (f628514; the gate ran on e78aad6, which differs by two cost files under tests/harness), in Sandboxes:

• Zero-cost gate over the whole corpus against main 5295319: 5,584 identical, 0 differ, 351 fail on both sides, 246 runtime-only (the programs that link the 'setImmediate code: the exception recorded for #3182), and the one island artifact (node-island-timers-typescript`: no libnode pack in the gate). Compile time: benchmarks -0.10% [-1.35%, +1.22%], self -0.16% [-2.37%, +1.03%], fixtures -0.69% [-1.87%, +0.49%].
• Island conformance lane, check mode: 442 passed, 0 failed. Island value matrix: 239 of 239. Weak-keys fixture: 3 of 3 plain and sanitized. Island harness tests: 225 plain, 22 sanitized. Unit tests 1,901, diagnostics 408. The four programs this train adds for the child and the working directory: plain and sanitized pass.
• 14 child-process island programs against Node, 50 runs each: 700 of 700 plain, 700 of 700 sanitized; ThreadSanitizer, 20 runs each: 280 clean. 13 worker island programs: 650 of 650 in each lane, 260 clean.

The full round with both native bootstraps is the one in the description (on 90063fc); since then the branch gained two clean merges of main and the two review changes, which touch only the island child unit and the island arm of four stream entries.

HOLD stays: do not merge yet. Four macOS jobs of this head are red (failed steps "Node islands (Intel)" and "Tests (darwin suite)") and their logs have not been read yet: they may be this train's new child-process or working-directory programs on macOS.

[cramforce]

Hold lifted on head f628514.

• Both review findings are answered and resolved: the pause / resume of an adopted child's stream was real and is fixed in a3419c9 (corpus 9896; the previous head differed from Node in 20 of 20 runs in each lane); the rest-parameter close listener did not reproduce (corpus 9895 equals Node on both heads, 20 of 20 in each lane), and the change now spares that listener explicitly.
• On this head: the zero-cost gate over the whole corpus shows no differing program (246 runtime-only rows under the recorded timer exception, IR and LLVM identical, -48 to +24 bytes); the island conformance lane passes 442 of 442 in check mode; the value matrix is 239 of 239 with no re-pin; 14 child-process and 13 worker island programs equal Node in 50 of 50 runs each, plain and sanitized; ThreadSanitizer is clean over 20 runs each. Both lanes and both native bootstraps passed on the tree before the last three merges of main and the review fixes.
• CI: lint and every Linux job are green. The four red macOS jobs fail one test that is red on main and on unrelated pull requests today (9811-node-island-load-reads-clock and its harness test node-island-load-reads.test.ts); it is not one of this train's programs.

🤖 Generated with Claude Code