AVP2 generated collector control, 7 November 2026
The required cancellation ordering reproduces in local Chromium even when a real fetch stream consumer receives every correct byte and observes EOF. It also reproduces for actual cancellation without EOF. Browser cancellation alone therefore cannot establish loss of the application's bytes. The existing collector rejects both. No safe source fix is justified by these indistinguishable collector-side observations without adding a consumer completion contract. This run preserves the discriminator and stops; it does not clear the original Marine gameplay prerequisite or establish historical AVP2 consumption.
Evidence: scratch/runs/20261107-avp2-collector-control/ (ignored). No guest,
disassembly, original replay, third-party browser internals, held-risk retry,
remote box, nested worker, or public deployment was used.
Exact consumed-byte witness
bound11-on-stream.json is the decisive real-browser trace. The unmodified
tools/browser-response-body-collector.cjs and the retained predecessor's
body-adapter.cjs served only generated HTML and a generated binary body.
Chrome 151.0.7922.108, Puppeteer 25.7.0, headless local Linux, same-origin
HTTP, cache disabled, no service worker, no WASM or game media.
The fixture is 57,344 bytes, each byte generated as
(i * 73 + (i >> 8) * 19) % 256; its SHA-256 is
509e350e05a572b1353fe40739100f48bec03145f1a5529d2a095ae051cd346e.
The consumer reads the genuine Response.body reader until done:true, checks
every byte against that formula, concatenates the actual returned chunks, and
hashes them with browser Web Crypto. It reports 57,344 bytes, zero mismatches,
the exact fixture hash, and EOF. There is no synthetic network event or callback.
For that one request the recorded order is:
| Sequence | Actual observation |
|---|---|
| 11 | HTTP200 response-stage body captured, full fixture hash |
| 13 | predecessor adapter policy verifies full body |
| 14 | Fetch.continueResponse completes; collector records continued |
| 15 | Network.responseReceived, HTTP200, type Fetch |
| 16 | Network.loadingFailed, canceled true, net::ERR_ABORTED |
| 17 | sticky collector network-loading-failed fault |
| 18 | real consumer returns all correct bytes and EOF |
Owning session BFF544CC2460A132AD67EE2F318B56CB, target
2655EBF2D47E857AD49DC592AF437CB7, network 1740771.2, Fetch request
interception-job-2.0. The independent consumer runs through Runtime.evaluate
on that exact session in default execution context3, unique ID
4832376845733028637.-2883291388823441718. Before/after target and document
identity match: http://127.0.0.1:42957/index.html, timeOrigin1794039815794.8.
The audit asserts matching target/session/network IDs throughout capture,
verification, continuation and cancellation, and matching consumer identity/hash.
Fetch's paused resource type is XHR while Network's request/response type is
Fetch, exactly as in the retained original failure.
Twelve paired controls using this bound consumer all read the correct bytes and
EOF: 12 interception-off and 12 interception-on. One intercepted pair produced
ERR_ABORTED; eleven did not. The earlier page-evaluate stream controls also
preserve two such cancellations among four intercepted runs. The samples show
intermittence, not a cancellation-rate or browser-internals explanation. The
initial arrayBuffer pair succeeds with interception on and off. Worker consumers
were unnecessary to reproduce this page-owned prerequisite and were not tested.
First concrete error and preserved negatives
The first incorrect inference would be treating the collector's sticky
network-loading-failed as proof that the fetch consumer did not receive its
body. Its loadingFailed handler faults after the narrow optional404 Script
exception; required200 has no consumer completion path. The adapter proves only
captured bytes and forwards this fault. drain() then prevents later capture.
The collector correctly reports the real CDP event; this experiment does not
justify deleting its cancellation guard.
bound-negative-*.json preserves the important counterexamples:
- Explicit AbortController abort: genuine AbortError, no completion, collector
rejects
ERR_ABORTED. - Read all57,344 correct bytes, then
reader.cancel()before adone:trueread: same hash andERR_ABORTED, no EOF, rejected. A hash alone is not an adequate completion discriminator. - Wrong body, empty/missing required body, and a truncated socket response: required-body policy rejects; no successful consumer completion is recorded. In the truncated case Chromium exposes a partial paused body, rejected by size/hash before continuation; the waiting consumer reaches its deadline.
- Detach the owning CDP session or replace its document after headers but before reading: no completion witness. The generated acceptance gate rejects both even though capture-only collector drain passes. Neither is retroactively called consumed.
on-headers-only.json: collector snapshot complete:true, consumer completion false. Collector completeness is a body-capture property, not consumer proof.
Every negative remains a failed acceptance in the generated controls. No
production cancellation, body coverage, identity, or timeout policy changed.
A comparison replacing the actual continuation command with
Fetch.continueRequest still produced the full-consumption cancellation. The
three small supplementary direct-protocol continuation controls happened to
succeed; they do not justify switching or replaying/fulfilling responses.
Concrete next step before original Marine gameplay
The missing prerequisite is an adapter receipt for actual consumer bytes plus EOF bound to the exact request, session and document. Retained original logs have none, so they cannot be reclassified using this generated result.
A follow-up must first add and test an opt-in consumer receipt contract: bind
one real application fetch to its Network request identity; observe its actual
reader/arrayBuffer completion without a second fetch; verify the resulting
byte length/hash against the captured pinned body; verify unchanged execution
context/document/session. A completion deadline, ambiguity, owner replacement,
missing EOF, body mismatch, real abort or uncovered required body must remain
failure. Only the matching completed consumer may distinguish its own
post-continuation cancellation. The old default must still reject unexplained
cancellation. The exact bound11 control is the old-failing regression; the
preserved negatives are the required opposing cases. Captured bytes, server
finish, continued state, URL alone or an application running flag are insufficient.
Only after that contract passes real Chromium controls, focused tests and a locked build should a separately authorized original registered Marine session be considered. This run launches none and makes no gameplay or emulator claim.
Reproduction, source, tests and lifecycle
Run from repository root:
PREFIX=fresh- STREAM_ONLY=1 node scratch/runs/20261107-avp2-collector-control/discriminators.cjs
PREFIX=fresh-neg- node scratch/runs/20261107-avp2-collector-control/discriminators.cjs
node scratch/runs/20261107-avp2-collector-control/audit.cjs
Use new prefixes; the first command is intermittent. The audit reads the retained
successful discriminator, checks exact identities/order and the consumer bytes,
and verifies rejection of the retained negative controls. It passes. Initial
controls and direct-protocol comparisons are also retained as small generated
inputs and logs. Source base is e83eb9fd315c24d525c1dd32b1f4181d974c33b1;
source receipts pin the actual collector, adapter, tests and generated controls.
No source tree or media was archived.
node test/test-browser-response-body-collector.js: PASS99 synthetic CDP
lifecycle cases with real local HTTP bytes; these are explicitly not Chromium
acceptance. node test/test-return-fire-body-driver.js: FAIL its preexisting
sealed ledger comparison for lib/browser-shell.js (expected
b7380c6dc115ab89794b491164fba0107788647314fdfe1a9276a1d7ae6624e9, actual
7d885b1b5a102493bc2f8f4fe9167ba62c271c09c5e09e87cfc787c67cebd5f1).
The exact failure is retained; no ledger or unrelated file was changed. No
collector/adapter/test source changed, and no new locked-build claim is made.
This is a documentation-only publication.
All generated local browsers and HTTP servers closed normally, their temporary
profiles were removed, and no remote lifecycle is owned. Historical
bx_z4d6hnmp was not accessed, resumed, stopped or deleted. Source changes to
tools/mcm-bound-input.cjs were avoided. Free disk stayed above2GB at checks;
the run uses approximately1MB. Artifacts precede result.json, written last.