[cramforce]

One rollup of three branches about the Node island boundary and one cleanup (landing train islands-2, first cut), merged together and validated as one tree. None of the three branches had a pull request of its own.

It removes the one answer #3377 (train islands-1) landed worse, and pays back part of what #3377 cost every dynamic program. #3377's first paragraph named it: structuredClone of a program error that went native code, island, native code gave fresh:{"%error":true,"message":"plain"} where Node gives fresh:{}, because structuredClone of a native error was already wrong with no island in the program. With this train structuredClone of an error gives an Error, as Node does, and the nine pinned n2i2n/U.clone = wrong cells and the four error control cells leave the island value matrix. #3377 also made 3,709 dynamic programs 320 to 1,064 bytes larger; 3,681 of them get 256 to 744 bytes back here (see "The island member sites" below).

What lands

The member branches are merged, not rebased, so their signed commits are in this branch.

• Branch fix/structured-clone-error (531b694b5c): structuredClone of an error in compiled code gives a new Error of the class Node picks (the one of Error's six native subclasses the source's name spells, else Error), with the source's own message, its stack, a clone of its cause and nothing else; a DOMException clones to a DOMException; one copy per source error within a call, so an error held twice clones once and a cause may lead back to its error. Before, the clone was a plain object carrying a hidden %error marker, a subclass's fields and an AggregateError's errors, without a class, a name or a stack. Two corpus programs (9917-structured-clone-errors.ts, 43 lines of output; 9918-structured-clone-errors-js.mjs, 19). Runtime only: the built compiler is byte-identical to main's. Loud as on main, and listed in the owner's notes: a static type that is a built-in or user Error subclass (SC1100 at build time, or the failed cast at run time), a cycle through a plain object, a cause that does not clone.
• Branch fix/island-matrix-wrong (6050e43a2b): a build-time warning, SC6007 "divergence: island value crosses as a copy", where the compiler sees a checked conversion copy an object or array a Node island made: an assertion islandCall() as T (also through a const or an await), a typed binding of such a value, a typed parameter of a callback passed to an island call when the callback writes through it or compares it, and an island call declared to return a Map or Set of native arrays or records. Since Node parity: compile the apps the island scorecard refused #3052 such a conversion copies where it used to throw, so identity is lost and writes are not shared, silently: 299 facts in 166 entries of the value matrix. The entries stay pinned wrong (the fix is a live view); the warning says where. Over the matrix's 239 programs: 44 sites in 20 programs, none in a control program; 298 of the 299 facts are behind a warned conversion. TypeScript and docs only: it changes no observation (the regenerated matrix is byte-identical with and without it). New limitations entry "Objects converted from Node islands", coverage fixture tests/coverage-fixtures/island-copies/.
• Branch fix/island-collector-leaf-records (cd09b5c6f9): the cross-heap collector releases a record or class instance without a cycle header that an island was handed. Such a record's capsule never had its reference counted by the walk before a major collection, so the record stayed a root for the rest of the run: 90,000 { attempt: number } records handed to an island left 90,000 exports held and two major collections freed none. 35 lines in scr_node_island_gc.c, fixture tests/fixtures/node-island/gc-records.{ts,cjs}, one test in node-island-gc.test.ts.

What the train changed beyond the members

• docs/content/docs/limitations.mdx: main (Deliver a child's stdout and stderr, and its children's events, in Node's order #3343, child process event order) and fix/island-matrix-wrong each added an include after the same line; both are kept.
• The island value matrix is regenerated on the train's tree by its update mode: sixteen pinned observations leave and one loses its clone=wrong part, nothing is new and nothing else changes. Leaving: U.clone = wrong in the four error control cells (error.control.-, error-cause.control.-, error-class.control.-, error-aggregate.control.-); n2i2n/U.clone = wrong in error.p, error.w, error.u, error-cause.w, error-cause.u, error-class.p, error-class.w, error-class.u, error-aggregate.u; clo-n2i/N.callback = clone=wrong in error.u, error-cause.u, error-aggregate.u. island-error-class.u clo-n2i/N.callback goes from clone=wrong,forin=wrong,json=wrong,keys=wrong to forin=wrong,json=wrong,keys=wrong. A full update of all 239 programs on the members' tree writes a file byte-identical to the committed one, and the matrix passes in check mode on the tree with main's islands: a value an island made passes the checked conversions #3029 and the cleanup.
• The CI cost files of the two new corpus programs.
• No bridge change (ABI stays 19), no new ScrNiDynOps entry, no change to the embedded host script or to the conformance list, so the train does not meet the start-up train (Train startup 1: start Node for island packages only when the program needs it, and answer island module resolution at build time #3350) in a generated file.
• Main is merged twice, without a conflict: up to 14960424a1 (Run the island conformance lane on pushes to main #3374, islands: a value an island made passes the checked conversions #3029) before the last gate round, and up to 793159ec8a (trains perf-2 and cli-2, Judge nursery bypass statistics per closure body #2719, Order a Node island lifecycle program by its steps, not by milliseconds (main is intermittently red in it) #3380, Bound computed import() specifiers and load bounded builtins natively; require in ES modules throws #2567, Fix main: remove an import binding-analysis.ts no longer uses #3381) for the push. The second merge brings 167 files; none is a file this train changes and none is an island runtime file.

The island member sites of scr_json.c do less (commit 1bda0dd984)

#3377 put five island paths of scr_json.c behind the island table (error_get, error_has, promise_get, promise_has, promise_set). Each site still tested its entry for NULL, and the two error sites looked the error up in the error-view cache before they called. That unit is in every dynamic program, so the tests and the lookup were too: #3377's gate showed 3,709 programs 320 to 1,064 bytes larger, and the lookup (scr_errdyn_find) had stopped being inlined.

The table is installed whole by one initializer, so an entry test decides nothing, and finding the error of a view is only needed in a program that links the island units. A site now tests the table and calls; error_get and error_has take the view and find its error in the island unit. The table's slots keep their places (only the first parameter of those two entries changes type).

function (release runtime object)	before #3377	#3377	this train
scr_dyn_in	1,748	1,982 (+234)	1,854 (+106)
scr_dyn_key_set	4,231	4,444 (+213)	4,302 (+71)
scr_dyn_obj_read_receiver	2,246	2,357 (+111)	2,251 (+5)
scr_dyn_promise_key_get	75	164 (+89)	154 (+79)
scr_errdyn_find as a function of its own	inlined	163	inlined

Each site is proven to still run by switching it off alone in one runtime build of that commit's tree; a named corpus program then stops matching Node, and with none switched off all pass: error_get and error_has: node-island-lifecycle-error-properties; promise_get: node-island-crossing-promise-subclass, node-island-crossing-request-promise and its typed twin; promise_has: node-island-crossing-promise-subclass, node-island-crossing-request-promise; promise_set: node-island-crossing-request-promise.

Validation

All in Vercel Sandboxes, Node 24.22.1, libnode pack built from the tree under test (bridge ABI 19). Three layers: one full round on the three members; then, for main's #3374 and #3029 and the cleanup commit, the gate with bench-runtime, the island suites and the mutation proof on that tree (1bda0dd984); then the short list on the pushed head, which adds only a merge of main.

Full round on the members (b55c3a4fb7 against main 3dda5d9c41's tree; bin/validate --gate-corpus=all with the allow lists below; 43.5 minutes, 93 lane jobs):

• Typecheck: pass (tsc in packages/cli and packages/compiler on a fresh tree).
• Zero-cost gate: PASS: 5,735 programs, 5,450 byte-identical, 229 fail to build on both sides (the island programs: the gate has no libnode pack), 56 allowed (the structuredClone rows below), none differing or runtime-only outside the lists. Compile time within tolerance: bench geomean +0.54% [-0.87%, +1.54%], fixtures -0.83% [-1.19%, -0.48%], self +0.03% [-4.08%, +3.37%] (inconclusive at the default run count).
• bench-runtime: 69 workloads, geomean -0.3%, 69 neutral.
• Lanes: plain 23,660 passed, 1 failed; sanitized 20,279 passed, 1 failed. The plain native bootstrap, the island harness suites, the island corpus and the value matrix in check mode are among the lane jobs and pass in both lanes. The two failures are timing, and neither program is touched by a member:
  • plain, llvm-differential (dev), node-island-lifecycle-first-use-in-handlers/main.mjs: the line island timer after uncaught came after unhandled second. The program (Node islands: fix error properties, process.argv, thenable awaits and work after 'exit'; pin lifecycle behaviour with 40 corpus programs #3123's, on main) starts Node inside an uncaughtException listener at 1 ms; the island's timer is due 1 ms after that and a native rejection at 15 ms, so a slow Node start flips the two. Alone, 20 runs of both differentials in one Sandbox: this train 20 passed; main's own tree 19 passed, 1 failed with the same output. Main has re-ordered the program since (Order a Node island lifecycle program by its steps, not by milliseconds (main is intermittently red in it) #3380, in this branch through the merge of main).
  • sanitized, server.test.ts https-request-fn: stopped at its 120 s limit. Alone, sanitized, 5 of 5 pass (63 s for the first build, under a second after).
• On the same tree in one Sandbox: pnpm node-compat:check current (4,996 source rows, 3,886 public rows); scripts/checker-identity.mjs --filter 991: 11 programs agree; island value matrix, full update: 239 of 239, byte-identical to the committed table; island conformance lane, check mode: 424 passed, 70 skipped, 0 failed; docs gate (NEXT_DIST_DIR=.next-check pnpm check): 25 of 25.

On the train's last own commit (1bda0dd984: the members, main 14960424a1, the cleanup commit; 32 vCPU, Intel Xeon 2.90 GHz):

• Build, runtime pack, generated-file checks (node-island-host.mjs, IR codec, IR fields): pass. pnpm lint: 0 warnings, 0 errors, the runtime refusal baseline included. pnpm format:check: clean. pnpm test:ts7 --baselines-only: complete.
• Guard tests (pack matrix with "a unit that tests SCR_NODE_ISLANDS has a build that defines it" and "the installed island table has every entry the shared units call through it", runtime define sites, builtin attributes, runtime features, executable cache, unit journal, peel guard, CI lanes): 185 passed. Island unit tests: 238 passed.
• Island harness suites, plain (every tests/harness/*island*.test.ts except the conformance lane; the value matrix in check mode and the new collector test among them): 383 of 383. Island corpus, plain (differential and llvm-differential, every island program): 470 of 470. Island diagnostics: 12 passed.
• Island conformance lane, check mode, 8 shards: 424 passed, 70 skipped, 0 failed. The list does not move.
• Sanitized: the island corpus with the cleanup commit on the members' tree passed 470 of 470 (and 470 of 470 plain); the sanitized harness suites and corpus on 1bda0dd984 are running beside CI, and I will add their result as a comment.
• Zero-cost gate with bench-runtime, --gate-corpus=all --gate-complete, against main 14960424a1: rows below.

On the pushed head (98b69349ad = 1bda0dd984 merged with main 793159ec8a; 16 vCPU, Intel Xeon 2.50 GHz): build, runtime pack, generated-file checks, pnpm lint (0 warnings, 0 errors), pnpm format:check, ts7 baselines, pnpm node-compat:check: pass. Guard tests: 227 passed, 1 failed, and that one is main's since #2567 (551639137b): outer-expressions.test.ts "only outer-expressions.ts defines a function that peels wrappers off an expression" names frontend/binding-writes.ts#stripValueWrappers, a file this train does not touch. Focused corpus (the two new programs, main's re-ordered lifecycle program, island-value-conversions, the mutation proof's programs; both differentials): 16 passed. The island tests of coverage.test.ts: 7 passed. Docs gate: 25 of 25. The gate's identity on the pushed head runs beside CI and goes into a comment.

Gate rows

Against main 14960424a1, 5,735 programs: 1,769 byte-identical, 229 fail to build on both sides, 56 allowed, 3,681 runtime-only, none differing. Compile time within tolerance (bench +0.57% [-0.87%, +1.82%], self -0.62% [-2.14%, +0.40%], fixtures +2.00% [+1.46%, +2.54%], which the tool calls inconclusive at the default run count; the members' own round read -0.83% for the same compiler). bench-runtime: 69 workloads, geomean +0.0%, 67 neutral, 1 faster (stream-buffered-drain -1.8%), 1 slower (csv-numbers +2.8% [+0.2%, +17.5%], whose executable is byte-identical to main's, so that reading is of one binary against itself).

• All 37 fully typed programs are byte-identical, and so are the 38 benchmarks that link none of the changed functions. No program gains a dynamic-tier or island symbol.
• The 3,681 runtime-only rows are every one smaller: 256 to 744 bytes (median 272; 1,832 programs -256, 803 -616, 674 -680), at most -0.20%. They are the cleanup commit: IR and LLVM identical on every row; the resized functions are scr_dyn_key_set (3,681 programs), scr_dyn_obj_read_receiver (3,680), scr_dyn_promise_key_get (2,024), scr_dyn_in (1,665), and scr_errdyn_find stops being a symbol in 1,664. 3,657 corpus programs and the 24 dynamic benchmarks. This is recorded as a one-time change in the fleet's gate exceptions, beside Train islands 1: errors, promises, process events and process.stdout.write keep Node's behaviour across the Node island boundary #3377's row, which it pays back in part.
• The 56 allowed rows are fix/structured-clone-error's, the programs that link the clone: 34 by the source pattern structuredClone (the two new corpus programs among them) and 22 by program id: performance.mark / measure clone their detail (6), JavaScript that reads builtin globals as values links the table that holds structuredClone as a function value (11), npm packages in the program call it (5). In the members' round, where they are the only rows: IR and LLVM identical on every one, the only resized functions scr_sc_clone and scr_structured_clone; 53 programs grow 2,768 to 3,000 bytes (median 2,936) and three that had no error-node code before grow 17,096 to 17,120 bytes (structured-clone-collections.ts +3.2%, 7614-key-order-copies.ts +3.1%, unknown-array-holes.ts +4.2%: the error node builder brings scr_dyn_prototype_unsupported and scr_dyn_prototype_valid), under the fleet's rule that a program using a rare feature may grow (--gate-max-size-growth=5).
• fix/island-matrix-wrong and fix/island-collector-leaf-records add no row: the first changes no emitted code, the second only an island unit.

Not run

• The full lanes were not repeated after the cleanup commit and main's later merges; they ran on the members' tree. The cleanup can only change what a program with islands does (in every other program the table is NULL and the sites skip), and the island suites, the gate and bench-runtime ran on its tree.
• The sanitized native bootstrap and Test262: no member changes what the compiler emits (fix/structured-clone-error leaves the built compiler byte-identical; fix/island-matrix-wrong adds a scan after lowering, and its own round on main 5b64e49735 ran both bootstraps and Test262 with nothing regressed). The plain native bootstrap is among the lane jobs.
• Darwin. pi.

Sequencing

The start-up train #3350 conflicts with island changes through two generated files (the embedded host header and the conformance list). This train touches neither, but by the fleet's sequencing it merges after #3350; if #3350 moves either list under it, main is merged and the generated island files are regenerated by their scripts before it lands.

🤖 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 3:42am UTC

[cramforce]

The two results the description promised.

Gate identity on the pushed head (98b69349ad against main 793159ec8a, --gate-corpus=all --gate-complete, 20 minutes): 5,756 programs; 1,772 byte-identical, 229 fail to build on both sides, 56 allowed (fix/structured-clone-error's rows, as in the description), 3,699 runtime-only, none differing. All 37 fully typed programs are byte-identical. The 3,699 runtime-only rows are the cleanup commit and every one is smaller: 256 to 744 bytes, median 256; IR and LLVM identical on each; resized scr_dyn_key_set (3,699), scr_dyn_obj_read_receiver (3,698), scr_dyn_promise_key_get (2,034), scr_dyn_in (1,674). The count is 18 above the description's 3,681 because main gained dynamic programs between the two bases. Compile time within tolerance (bench -0.36% [-0.77%, +0.38%], self +0.17% [-1.02%, +1.64%]); the candidate compiler source builds with the candidate.

Sanitized island suites on 1bda0dd984 (the members, main 14960424a1, the cleanup commit; 32 vCPU, libnode pack from the tree): island harness suites 383 of 383, island corpus (differential and llvm-differential, every island program) 470 of 470.

Main is red in one unit test since #2567 (551639137b), and it will show in this pull request's plain lane: packages/compiler/src/frontend/outer-expressions.test.ts, "only outer-expressions.ts defines a function that peels wrappers off an expression", names frontend/binding-writes.ts#stripValueWrappers. This branch does not touch that file.

[cramforce]

Pushed 19cd362602, one commit on the head: the review finding about an accessor-only cause (thread above; reproduced against Node, fixed, pinned in 9918-structured-clone-errors-js.mjs). The second finding (a deleted stack) is answered in its thread: the plain forms are named refusals, and in the one form that compiles the difference is in the delete itself, on main as well; it is recorded as a follow-up.

The commit changes one function, scr_sc_clone_error, which only the programs that link the clone carry (the 56 allowed rows of the description), and one corpus program. On it, in a Sandbox (32 vCPU): build, generated-file checks, lint (0 warnings, 0 errors), format, ts7 baselines; the clone programs in both differentials, plain 28 of 28 and sanitized 28 of 28; the island value matrix's error cells in check mode, 18 of 18; scripts/checker-identity.mjs --filter 991, 11 programs agree. Main was not merged again (a trial merge with 659a1caaa5 has no conflict).

[cramforce]

Pushed 0f4df1865e: main is merged up to 851b59b90f (trains cli-4, cli-5 and startup-1), and one snapshot is regenerated. This is the push the description's "Sequencing" section announced.

The merges.

• With train cli-5 (Train cli 5: a project's tsconfig decides import elision, its own declarations, and how the checker resolves them #3386), two conflicts, both in the diagnostics tables: its SC6008 and this train's SC6007 were added on the same line after SC6004 (codes.ts, and the family list in the header of diagnostic.ts). Both are kept.
• With train startup-1 (Train startup 1: start Node for island packages only when the program needs it, and answer island module resolution at build time #3350) and cli-4 (Train cli 4: compiler crashes get a source position, three of them fixed #3384): no conflict. The files both sides changed (diagnostic.ts, lower-node-island.ts, scr_error.h, scr_node_island_handle.c, scr_node_island.h) have independent hunks.
• tests/harness/__snapshots__/coverage-island-copies.txt is regenerated by its test's update mode: the fixture's statements in the dynamic tier go from 12 to 10 on main with startup-1. The SC6007 warnings it pins are unchanged.

The generated island files, each by its script on the merged tree (32 vCPU, libnode pack from the tree, bridge ABI 19):

• Embedded host header: node scripts/node-island-host.mjs --check is current (this train does not touch the host script).
• Island value matrix: a full update of all 239 programs writes a file byte-identical to the committed one. Against main's table the difference is this train's 17 entries, as in the description (main's cli-5 removed 29 .ctor refusals of its own in between).
• Island conformance list: the lane in check mode, 8 shards, 434 passed, 0 failed, so the list on main is exact for this tree and nothing is regenerated.

Checks on the merged tree. Build, runtime pack, generated-file checks, pnpm lint (0 warnings, 0 errors; runtime refusal sites 390, none outside the baseline), pnpm format:check, ts7 baselines, pnpm node-compat:check: pass. Guard and unit tests (pack matrix, runtime define sites, builtin attributes, runtime features, executable cache, unit journal, lowerer, peel guard, CI lanes, diagnostics, node-islands): 523 passed. Island unit tests 321. Island harness suites, plain: 423 of 423. Island diagnostics 12. Island corpus, plain (differential and llvm-differential, every island program): 578 of 578. The island and type-claim tests of coverage.test.ts: 12 passed with the regenerated snapshot. Docs gate: 25 of 25.

Native bootstraps. CI runs them only for pull requests that touch the bootstrap paths, and no round of this train ran both on a tree that differs from the head only by main: the plain and the sanitized native bootstrap (self-hosting-native-driver.test.ts, run by the validation driver's lane tool as a round runs them) are running on this head's tree in two Sandboxes; I add their result here. The sanitized island suites on this tree are running too.

[cramforce]

The results the last comment said were running, all on the pushed head's tree (0f4df1865e: main 851b59b90f, this train, the regenerated snapshot).

Both native bootstraps pass. tests/harness/self-hosting-native-driver.test.ts, run by the validation driver's lane tool exactly as a round's native lane job runs it, each alone in its own Sandbox:

• plain (16 vCPU, 32 GB): passed in 2,196 s;
• sanitized (32 vCPU, 64 GB, all phases in one job): passed in 3,232 s (the seed, the seed's command and library probes, the sanitized native emitter rebuilding the compiler, the sanitized native frontend lowering the compiler, Node's and the native compiler's LLVM compared).

Sanitized island suites (32 vCPU, libnode pack from the tree): island harness suites 423 of 423, island corpus (differential and llvm-differential, every island program) 578 of 578.