[vincentkoc]

This relates to...

Fixes #6110.

Rationale

Aborting an unread data response reaches an HTTP-only stream controller and throws instead of rejecting the body. Blob and integrity-replaced response bodies also need a body-owned error capability. Cloning must keep the fetched Response's current branch abortable without retaining the public Response.

Changes

• Capture stream error ownership where response bodies are constructed and attach the Fetch abort owner after integrity processing.
• Rebind that owner to the original branch when cloning; preserve the ordinary un-cloned HTTP path and producer-completion timing.
• Cover unread/default/custom abort, BYOB, completed-producer clone behavior, abandoned Responses, live non-HTTP clones after GC, reader retirement, source errors, and HTTP integrity replacement.

Features

N/A.

Bug Fixes

Preserve abort rejection and cleanup for non-HTTP and replaced Fetch response bodies.

Breaking Changes and Deprecations

None intended. Already-completed clone branches remain readable after abort.

Status

• I have read and agreed to the Developer's Certificate of Origin
• Tested
• Benchmarked (optional)
• Documented
• Review ready
• In review
• Merge ready

Local validation on the exact candidate:

• npm run lint: passed on Node 26.12.1.
• npm run test:fetch on Node 24.16.0: passed the complete native chain—Fetch 518/518, WebIDL 28/28, Busboy 49/49, with no failed, skipped or cancelled cases.
• node --expose-gc --test --test-concurrency=1 --test-timeout=10000 test/fetch/issue-6110.js test/fetch/issue-4799.js: 33/33 passed on Node 22.23.2 and 24.16.0, with no skipped or cancelled cases.
• Earlier bounded Node 26.12.1 commands exercised 33 distinct cases; the original unread-data regression failed before the production change.
• npm run test:fetch on Node 26.12.1: 518 tests, 517 passed, one failed: test/fetch/referrrer-policy.js, “should set referrer policy from response headers on redirect” (plan expected 1 assertions but received 0). The same file subsequently passed 12/12 on both the unchanged base and candidate under matched runtime/environment/runner settings. That isolated result does not explain or supersede the full-run failure.
• The first full-fetch attempt also encountered Node’s warning about conflicting inherited NO_COLOR/FORCE_COLOR settings in bundle-parser child stderr. Removing FORCE_COLOR for validation commands resolved that exact surface (8/8 passed), without changing source or assertions. The 517/518 result above used this corrected environment.

The test:fetch command stopped at the Fetch stage. Its sibling scripts were then run independently on Node 26.12.1: npm run test:webidl -- --concurrency 1 passed 28/28; npm run test:busboy -- --concurrency 1 passed 49/49 (no skipped or cancelled cases). One temporary observational run of the complete 12-case referrer file also passed: every case reached its own target assertion before fetch fulfillment and teardown. The instrumentation was restored afterward. This did not reproduce or explain the full-run failure; no source or assertion change was made for it. Full-suite/WPT, cross-OS, shared-builtin and performance qualification remain incomplete; the unchanged shared-builtin CI job is disabled upstream. Existing upstream CI, Benchmarks and CodeQL runs await maintainer approval and have executed no jobs. No personal-repository workflow runs were dispatched or executed.

Independent owner/cleanup source review and a fresh complete-source autoreview found no actionable findings. Ready for maintainer review; merge readiness and complete test qualification remain open because of the unresolved Node 26 full-run failure and pending upstream matrix/performance results. The source reviews do not replace runtime or performance qualification.

Candidate: e6c2b7f806c38ee47243a6286ecfaf7c2fea1ed2; tree 0a6fad486120bbb4a4f8158fd0ff9312771df2b3; base 609c6efca79f1592954d5fe10e58f69cda342a07. Workflows and package metadata are unchanged.

[vincentkoc]

Additional downstream evidence from OpenClaw's native qualification run:

• Job: https://github.com/openclaw/openclaw/actions/runs/41086168327/job/124168912430
• Source: ab6f13b2a2e215062f3a32cbd2f28c86d6f3b3bf; runtime logged as Node.js 24.22.1.
• All 733 files and 8,021 test assertions passed, but Vitest reported one unhandled error while running chat-attachments.test.tsx:

TypeError: Cannot read properties of undefined (reading 'error')
abortFetch node:internal/deps/undici/undici:13198:20
AbortSignal.<anonymous> node:internal/deps/undici/undici:13112:11
AbortController.abort node:internal/abort_controller:580:5
AbortSignal.abort node:internal/deps/undici/undici:12115:14

The official Node v24.22.1 bundle matches these logged frames: line 13198 dereferences controller.error(error), reached through line 13112 passing the HTTP-owned controller. Its bundled Undici package is 7.30.1; the CI log did not explicitly print that dependency version. This identifies the same ownership boundary addressed here, while the exact aborting request remains unproven. The unchanged downstream text-card fixture performs a real data-URL Fetch and aborts during disposal, but the log captures neither the request URL nor its precise test phase. This source evidence does not establish that this patch makes the historical job pass. The existing CI, CodeQL and benchmark runs for e6c2b7f806c38ee47243a6286ecfaf7c2fea1ed2 currently require maintainer Actions approval. No downstream dependency override or new fork was created.