node: errors/whatwg/promise v26 compat fixes (+11 tests)

[cirospaciari]

Vendors 11 previously-failing Node v26.3.0 upstream tests (byte-verbatim) and fixes the runtime/harness gaps they found. Stacked on claude/callback-throw-uncaught.

Runtime fixes

process promise-rejection plumbing (BunProcess.cpp, ProcessObjectInternals.ts)

• 'unhandledRejection' listeners were invoked straight from native code with zero JS frames on the stack, so Error.captureStackTrace(err, listener) inside a listener produced an empty CallSite array (stack[0] undefined — this crashes node's own common.mustNotCall when used as a listener). Dispatch now goes through a tiny JS trampoline that forwards to process.emit, so listeners see caller frames like node's processPromiseRejections path. The fixing line is the processObjectInternalsEmitUnhandledRejectionFromNativeCodeGenerator call in Bun__handleUnhandledRejection; listener-exception handling is unchanged (same emitter).
• Warnings created from native context (no JS frames) had warning.stack === undefined; node always produces at least the "<name>: <message>" header. Process::emitWarning now writes that header when no stack was captured.
• PromiseRejectionHandledWarning was emitted even when a 'rejectionHandled' listener existed; node only warns when nothing handled the event (if (!process.emit('rejectionHandled', ...)) emitWarning(...)). Reordered to match.

[nodejs.util.inspect.custom] function name (WebStreamsInspectCustom.cpp)

• The custom-inspect method installed on URL/URLSearchParams (and web-stream) prototypes was an anonymous function; node names it [nodejs.util.inspect.custom] and test-whatwg-url-properties reads fn.name. The function is now created with that explicit name; property attributes unchanged.

internalBinding test shim (src/js/internal/test/binding.ts)

• New config (hasIntl), encoding_binding (encodeInto/encodeUtf8String/decodeUTF8 backed by the runtime's own encoders) and util.privateSymbols (arrow_message_private_symbol) entries, needed at module-evaluation time by the vendored internals below. New switch cases only — no overlap with the http2 (http2: implement internalBinding('http2') for the vendored node test shim (+2 upstream tests) #34499) or cares_wrap (tls: close_notify on end(), injected-socket upgrades, reject-handshake wire fix, duplex data-loss, SNI, ALPN (+14 tests, tls 81%→86%) #34598) cases pending in other PRs.

Test-harness fixes (test/js/node/test/common)

• Flags scan bug: the --expose-gc / --expose-externalize-string branches ended the scan with break, so // Flags: --expose-gc --expose-internals never installed the expose-internals require interceptor ("Cannot find module 'internal/...'"). They now continue like every other handled flag. The five base tests whose Flags lines are affected (test-fs-filehandle, test-performance-eventloopdelay, test-crypto-dh-leak, test-trace-events-api, test-v8-flag-pool-size-0) still pass.
• Vendored internals: adds byte-identical copies of node v26.3.0 lib/internal/errors.js, lib/internal/encoding.js, lib/internal/encoding/{single-byte,util}.js under common/nodeinternals/, plus the emulator entries they need (AggregateError primordial, uncurried %TypedArray% statics, a faithful FastBuffer shim, customInspectSymbol). The errors-family tests now run against node's real E/SystemError/AbortError/codes implementation.

Vendored tests (all byte-verbatim, all passing)

• test-errors-aborterror
• test-errors-systemerror-frozen-intrinsics
• test-errors-systemerror-stackTraceLimit-{custom-setter,deleted,deleted-and-Error-sealed,has-only-a-getter,not-writable}
• test-whatwg-url-properties
• test-whatwg-encoding-custom-internals
• test-promise-unhandled-error-with-reading-file
• test-promise-unhandled-warn

Each fails on the unfixed build (module-resolution error, fn.name === 'anonymous', or the stack[0].getFileName crash respectively) and passes with this branch.

Investigated and deliberately not vendored (root causes)

• test-http-raw-headers: request trailers are lost when the handler calls res.end() before the request body finishes — Bun emits req 'end' at dump time, before the parser has consumed the trailer section (repro in PR thread; overlaps http/http2: node v26.3.0 compat — HTTP/1 fallback + upgrade handoff, http2 session errors, perf_hooks and frame framing (+11 upstream tests) #34432's parser-spill rework, so not fixed here).
• test-errors-systemerror: asserts V8's exact TypeError text ("Cannot read properties of undefined"); JSC's differs.
• test-errors-hide-stack-frames: needs Bun's native validators to hide their own frame from e.stack (node's hideStackFrames contract) — wide-blast-radius change, tracked separately.
• test-abort-controller-any-timeout: only blocker is the TimeoutError message text ("The operation timed out." — WebKit — vs node's "The operation was aborted due to timeout"); an existing Bun test pins the WebKit wording, so changing it needs a maintainer call.
• test-whatwg-url-custom-searchparams-constructor: node deviates from WebIDL (init == null treated as absent); Bun follows WebKit/spec ('null=').
• http2 *-errors family, -binding, -ping, -debug, -padding-aligned, -socket-proxy, util tests: require node's native-handle architecture (mockable Http2Session/Http2Stream prototypes the JS layer routes through), async_hooks HTTP2PING resources, or node timer internals. (test-http2-binding/test-http2-respond-errors are already covered by http2: implement internalBinding('http2') for the vendored node test shim (+2 upstream tests) #34499.)
• webstreams transfer/internal tests: need node's kTransfer worker-transfer protocol or internal/webstreams/* implementation internals; -adapters-to-streamreadable conflicts with Bun's deliberate lazy-lock design (documented in node:webstreams: +17 tests from Node v26.3.0 (30% → 76% of the suite), fix the gaps they found #32627).

[cirospaciari]

🤖 Repro for the test-http-raw-headers root cause noted in the description (request trailers lost when the handler responds before the request body completes). Node prints the trailer pair; Bun prints []:

const http = require('http');
const server = http.createServer((req, res) => {
  req.on('end', () => { console.log('rawTrailers:', JSON.stringify(req.rawTrailers)); server.close(); });
  req.resume();
  res.end('x f o o');   // respond immediately, before the request body finishes
});
server.listen(0, () => {
  const req = http.request({ port: server.address().port, path: '/' });
  req.addTrailers([['x-bAr', 'yOyOyOy']]);
  req.setHeader('transfer-ENCODING', 'CHUNKED');
  req.end('y b a r');
  req.on('response', r => r.resume());
});

Moving res.end() into the request 'end' handler makes Bun match Node, so the loss is specific to the respond-early path: markRequestAsDone() runs at res.end() and JS emits req 'end' at dump time, before the parser has consumed the trailer section, so takeRequestTrailers() finds nothing. Also reproduces on the #33191 CI binary. Fixing it means deferring the dumped request's 'end' until native message completion — that sits inside the parser-spill machinery #34432 is reworking, so it is left out of this PR.

[github-actions]

This PR may be a duplicate of:

1. process: print Node-style warning report and hide internal frames in emitWarning #34608 - Both modify BunProcess.cpp's Process::emitWarning to make warning output match Node.js behavior when no user JS frames are present
2. util.inspect: install [nodejs.util.inspect.custom] on web-platform prototypes #35292 - Both modify WebStreamsInspectCustom.cpp to fix [nodejs.util.inspect.custom] naming/behavior on web-platform prototypes

🤖 Generated with Claude Code

[cirospaciari]

🤖 Coalesced into #35483 (same base, per review-load consolidation) — all 5 commits merged with a verified strict-superset check; closing.