AoW II current collector: owning-worklet transport qualified; original startup blocked

Run: scratch/runs/20261106-aow2-current-collector-terrain/ (ignored). Worker: aow2-current-collector1431, 2026-11-06.

The current browser-body collector now permits its existing, exact capability-checked owning Network transport on a paused worklet, as well as a dedicated Worker. Real Chrome establishes that the audio worklet lacks Fetch.enable, but emits its own request/response/finished chain and returns its own script bytes. This is an observation transport repair, not an emulator semantic or terrain repair. The special dedicated-Worker startup binding is unchanged; worklets must provide their own ordinary request identity.

The original did not instantiate the guest. Terrain creation/upload/draw contribution and current army gameplay remain unverified. Accepted earlier army displacement and movement-point reductions remain valid and untouched. No black-terrain observation here establishes an emulator defect.

Adaptation and controls

The retained AoW driver used page response.buffer() and lost request identity in its outer catch. The run-local adapter uses the current collector, following the actual Return Fire composition's paused-target attachment, transport setup, policy propagation, resume, cancellation and shutdown order. Worklets also lack Target.setAutoAttach; only its exact method-not-found error on a worklet is accepted by this composition. Worker/nested-Worker attachment remains required.

All HTTP 200 bodies must match pinned bytes. Optional 404s are restricted to explicit local static absences and the reviewed shell/font-bundle/DLL lookup paths; font-bundle fallback additionally requires five verified individual font bodies. Unknown failures remain fatal. Server hashes supplement browser receipts and never replace them. The retained per-Worker durable drain store, separate CPU/API/DX/GPU order accounting and observer quotas are unchanged.

Validation before the registered Launch:

Failed generated controls remain separately retained: unsupported worklet Fetch; an incorrectly proposed parent-only worklet transport, disproved by its own Network events; missing worklet Target domain; an assertion expecting only two receipts when parent Fetch and Worker both report transfers; duplicate detach commands; and awaiting success after deliberately holding a required404. None is relabeled as passing. transport-controls.json selects the seven passing proofs and seals current collector/adapter hashes.

Original boundary, including failed shell setup

The coordinator granted runtime at 15:02:12 UTC, deadline 15:47:12 including cleanup, after KKnD1321 and CompiledSparseSave1354 FINAL releases. Fresh independent checks covered 25+7 absent PIDs, three refused ports, two absent profiles and four initially free locks. No boxes or nested workers.

The first Chrome shell setup, 15:13:14–15:14:15, stopped before launchApp(): the pin-list constructor omitted tracked manifest.webmanifest. Its captured 598 bytes/hash matched both Git and the repository exactly. The driver had already closed Chrome normally on navigation timeout before repair; no Launch receipt or instantiated guest existed. All first-attempt artifacts remain in shell-setup-attempt1/. setup-pin-repair.json documents the exact omission; the pin-list recipe now includes .webmanifest. This is not a claim that only one Chrome setup session ran.

The next browser's sole registered Launch occurred at 15:15:16.921 UTC. At 15:15:16.965, the page-owning Fetch session received an error response for fonts/w95fa.woff2: responseErrorReason: Failed, with no HTTP status or body. fault-00001.json fixes URL/method/request/network/target/session identity. No completed server response for that font is recorded. The collector preserved this first fault and refused to continue later required responses. No unavailable body was ignored, retried through another channel or replaced by server bytes.

I personally reviewed initial.png: it shows Preparing emulator, not the AoW menu/world. Its state has running=false, guestWorker=false, no render endpoint or instantiated guest module. No gameplay input followed Launch. The main-module guard closed Chrome normally at 15:16:27. A subsequent repair connection found its DevTools port already refused; no intact guest remained. There was no further original retry. Original instantiated-module equality and owning WebGL are therefore not established in this run.

A generated font control subsequently loads the same font successfully with and without the same collector on the isolated static server. It preserves the font's actual browser body/hash and normal exit; it does not explain the original failure. Cancellation, cache, network and other causes remain unknown. The original yields zero CPU/DX/GPU drain events and zero terrain births; zero events do not imply missing guest drawing operations.

Provenance and remaining work

Tested main is ab6efb9f9748189f14a93b18ed5edda2c0f04c5a, with the recorded collector regression diff. All 301 guest/runtime source identities, both modules and 1,064 media path/size/hash references match before/after. Standard module: 43a0097ab79e1e9102d6f12411485a359ebe10b03af5fcf22675b75ad701a081; compat: 7219c6f2d7bcf6f91975a8b3bb9fbac35cbc5933c8fc47fe5d4a44b31f1c7a46. Those are controls/build identities, not a claim of original instantiation. The existing AoW2.exe remains SHA256 a10590e5dbd013d154b00ea53e66670f4e74d38ab33adb2523f0a662af8f89f7. Git SHA, binary diffs, helper patches and retained-input hashes provide recovery; no source-tree archive or repeated media copy was made.

The blocker is the original font request's failed browser response, compounded by automatic guard closure before same-browser repair. A future assignment must qualify recovery/observation for that concrete failure while preserving first fault and required-body/root identity. Terrain lifetime and two reviewed army moves with visible point decrement remain outstanding. This is an honest transport/startup partial result, not parity or full27 completion. No emulator, guest, media, input, viewport or vblank semantics changed. Box ID/state/expiry are not applicable; runtime ownership returns to the coordinator after cleanup.