Return Fire generated stream cancellation source — 2026-11-05

Generated full-byte cancellations reach Chromium's native ResponseBodyLoader::DidCancelLoadingBody before page/session teardown. Retained-reference controls do so while the response, reader and un-aborted controller remain live. This locates a browser renderer body-loader cancellation boundary. A directed consumer-timing control reaches DidFinishLoadingBody instead. No project collector/host lifecycle defect is demonstrated; no production repair, cancellation exception, completeness relaxation or transport change was made.

The likely internal mechanism is normal stream closure calling cancellation before the delegating loader records its terminal state. That call sequence is an inference from matching-version source, not a sampled C++ stack or a verified source-to-binary mapping. The original ReturnFire failure still lacks a consumed-byte witness and native trace; its cause remains unproved.

Evidence: scratch/runs/20261105-return-fire-stream-cancellation-source/. Requested input main: 9aab8b88068809f8d348fb31fde8f611b16287e0. Tested main: b5118b7bce827976e7c36bb93af981837e2cc13f. Relevant collector/driver/tests/host/byte-provider input and worktree diffs are recorded, along with exact file hashes and owned generated source patches. No source tree archive, media copy or module copy was made.

Directed generated experiment

The retained predecessor fixture supplies the exact consumer function, generated byte formula and production collector path, tools/browser-response-body-collector.cjs. Only its case 3 and case 21 shapes were reused. One headful browser lifetime ran seven directed cases in 13.921 seconds; no original game was launched.

Every successful consumer independently checked all 1,932,903 response bytes and WebCrypto SHA-256 19346dc34617036349024d6e4c273a1f8cc0bfd43119640369953ef52e555619. There were 16 full-byte consumers. CDP/trace/NetLog bytes never entered the consumer. Native tracing and default-mode NetLog only recorded observations; neither supplied response bytes nor controlled continuation or completion. The explicit timing control, rather than tracing, changes consumer scheduling.

Control Independent consumer Required Network terminal Native body-loader callback
Case 3 shape, unchanged consumer, no interception Full bytes/hash Failed, canceled ERR_ABORTED DidCancelLoadingBody
Case 21 shape, six collector streams Six full bytes/hashes Six loadingFinished Six DidFinishLoadingBody
Single stream, response/reader/controller retained Full bytes/hash; closed/done; signal false Failed, canceled ERR_ABORTED DidCancelLoadingBody
Six collector streams, references retained Six full bytes/hashes Six loadingFinished Six DidFinishLoadingBody
Slow stream, eager reads, references retained Full bytes/hash; closed/done; signal false Failed, canceled ERR_ABORTED DidCancelLoadingBody
Same slow stream, pause 100 ms before tail read Full bytes/hash; closed/done; signal false loadingFinished DidFinishLoadingBody
Actual AbortController after first chunk AbortError, 16,384 bytes; closed rejects Failed, canceled ERR_ABORTED No DidCancelLoadingBody event

The slow pair uses the predecessor's existing 16,384-byte/10ms generated server path, identical body/framing and retained references. The pause starts after 1,867,776 bytes, leaving 65,127 bytes. No reader.cancel, abort, forced GC, method replacement or production transport edit occurs in either positive control. The native callback changes with this scheduling distinction. This is one directed comparison, not proof of determinism across repeated runs.

The two collector cases captured/verified/continued all twelve asset bodies and were complete before cleanup. They did not reproduce the predecessor's collector cancellation in this run. This does not prove interception has no influence. All three positive cancellations here occur without Fetch enable, CDP body reads or a collector. The actual abort is a fourth failed transfer. The strict terminal checker exits 1 and requiredTransfersPass remains false.

Exact lifetime evidence

Retained single stream: owning session 2781C7D88A1C0C1A2AC12AF3677225BC, target D81BAECA8782E34FF4FCAD4CF7451469, request 3095769.2, renderer PID 3095769. Its request-ID-bound native ResourceFinish is failed. There is only one generated asset in that renderer's request window; its native callback is DidCancelLoadingBody, not DidFinishLoadingBody.

UTC, 2026-11-05 Observation
16:40:53.851100 Controller created; fetch called
16:40:53.857000 Server response finish and close, writableFinished true
16:40:53.863287 Native DidCancelLoadingBody
16:40:53.864153 Network loadingFailed, canceled ERR_ABORTED
16:40:53.865200 reader.closed fulfillment callback
16:40:53.865300 Actual reader.read returns done=true after 1,932,903 bytes
16:40:53.874200 Consumer verifies every byte/hash; signal aborted=false
16:40:54.229900 Explicit reference release
16:40:54.333000 Owner calls page.close
16:40:54.343000 Owning session disconnect observed

Page times use performance.timeOrigin+now; native times use NetLog's recorded TimeTicks offset. Native microsecond fields are retained, but cross-clock wall conversions are not a claim of microsecond measurement accuracy. Network loadingFinished can carry upstream finish time earlier than renderer emission; native correlation therefore uses the exact request-ID ResourceFinish trace window, not a loadingFinished timestamp as an emission-time bound.

Before teardown, every case queries Target.getTargetInfo through its original owning session and confirms the same target and an open page. Snapshots are persisted before collector.close, reference release and page.close. The first failure is preserved separately as baseline request 3095705.2; no session reattachment, parent substitution or cleanup retry was used to reinterpret it.

For each full-byte cancellation, the exact URL_REQUEST NetLog source reports 1,932,903 filtered bytes and no CANCELLED/FAILED event. Server finish/close is normal. The actual abort has a NetLog CANCELLED event and partial bytes. These counts corroborate transport delivery; they do not replace the independent JavaScript byte/content witness. NetLog is diagnostic logging, not a production completion API, as its primary design document explains.

Source-supported mechanism and causal limit

The 151.0.7922.108 body-loader source notifies its downstream client in OnStateChange before recording the terminal public state. Cancel tests its internal loading state and can select CancelSync/DidCancelLoadingBody. In the body-stream source, normal closed-state handling calls Close and then CancelConsumer. The resource loader turns DidCancelLoadingBody into Cancel/CancelledError. This explains how a normally closed consumer stream can coexist with a failed Network terminal, without any explicit JavaScript abort or teardown. Native traces and the tail-timing comparison support this explanation but do not expose the internal state values or the complete C++ stack.

Live Browser.getVersion reports Chrome/151.0.7922.108, V8 15.1.206.16 and a zero source revision. Installed dpkg metadata says 151.0.7922.71-1. Actual /opt/google/chrome/chrome SHA-256 is 22e4d5ef0c637a87b59a9baad42625aba3284dc03d4a5908748e253556318c97. The source tag is therefore a candidate implementation, not verified binary provenance. Primary URL/file hashes and relevant excerpts are retained. A Gitiles history request returned HTTP 403 at 16:20:08Z; that history operation stopped without retry or alternate-history access. No upstream-fix claim is made.

The Fetch garbage-collection rules alone do not identify this cancellation. It occurs with strongly retained response/reader/controller, normal closed/done observations and no signal abort, well before explicit lifetime cleanup. The direct demonstrated boundary is browser body-loader cancellation; the exact bookkeeping/reentrancy sequence remains an inference. Next discriminating evidence would be a source-mapped native stack/state capture at DelegatingBytesConsumer::Cancel/CancelSync, BodyStreamBuffer::Close and HandleResult on this same generated fixture, showing terminal notification order. No browser/version sweep or speculative project change is warranted.

The original immutable failure remains SHA-256 bfc438fad8a6173442ba8867f276bc948a4812de2f7b5093bebd24ab15ee5f78. Neither full original delivery nor this native path in that original is proved.

Validation and cleanup

The unchanged collector suite passed 91 cases; the unchanged composed driver suite passed all 33 synthetic CDP/real-FSHTTP cases in 134.977 seconds, with zero Chrome launches. Generated causal assertions passed while explicitly retaining four failed required transfers. Scratch analysis corrections joined server path/query to full URL, selected NetLog URL_REQUEST source identity rather than its separate HTTP-stream controller, and bounded renderer events by ResourceFinish. Earlier checker outputs/source patches remain; no terminal failure was suppressed. flock /tmp/wine-build.lock bash tools/build.sh passed in 30.121 seconds. Normal Browser.close exited PID 3095627, startticks 194406078, code 0/null; no kill fallback ran. Cleanup verified 31 recorded browser/command identities absent, both HTTP/DevTools ports refusing, the own profile removed, four locks free and no heavy runtime. The own board watcher is separately stopped and verified in the final cleanup artifacts.

Runtime began only after Lords1542 FINAL16:25:24Z, coordinator Strong1555 recovery FINAL16:28:38Z, then rectangular1610 FINAL16:40:09Z, all with runtime and file release/ALLBUILDSFINISHED. Independent checks found 241 predecessor PIDs absent, 98 ports refusing, 22 profiles absent, four locks free and over 8 GB disk free. The held Stronghold lane was not retried. No test box was assigned: ID/state/expiry N/A; lifecycle ownership remains with the coordinator. No public deployment or original launch occurred.