Recoil Software browser source transport repair, 2026-11-08

Run: scratch/runs/20261108-recoil-software-browser/.

Completed the first new tested generic integration repair alternative: explicit UTF-8 transport for generated browser source modules and their HTML entry page. The single authorized local Chrome attempt exposed this defect in the real [[FunctionLocation]] / Debugger.getScriptSource path. The corrected transport has not been retried in Chrome. Software ownership and its explicit gate remain unvalidated; no gameplay consequence is claimed.

Real attempt and exact cause

Read the predecessor note at 807e003e2 (the task's concatenated recoil-software-owner-20261108.md807e003e2 resolves to that file and commit). Copied its corrected browser-integration.cjs, local-supervisor.py and repaired absent-renderer bound-input.cjs into this new run. Independently verified all 851 provider pins and retained the source/startup/motion closure. The accepted 28 focused negatives, exact production-copy test, absent-renderer regression and locked full build were reused by exact log hashes in tests-reused.json; production code was unchanged. Accepted point/custodian controls were also reused by verified manifest hashes.

Before Chrome, PREFLIGHT_ONLY=1 passed, all 41 exact generated module wrappers passed vm.Script syntax checks, and the fixture retained host.dx_trace(21,7,2,0,0). An initial concern about wrapper newline escaping was disproved by those checks; no wrapper syntax defect is claimed.

Exactly one generated, local Chrome attempt ran. It reached an actual software copy and passed the independent same-host clock bracket 1794146019781–1794146019783, target1794146019783. The decoder source response was correctly refused because its hash differed from the pinned UTF-8 source:

The server sent text/javascript without an encoding and served an HTML page without an encoding declaration. Decoding the exact decoder wrapper's UTF-8 bytes as Windows-1252 reproduces Chrome's observed hash exactly. This establishes a concrete transport failure rather than an inferred owner change. The retained failure and attempted script remain unchanged. The attestor's parallel lanes drained; no successful source proof was recorded. Response bodies were not saved by that attempted fixture, so this run does not claim four complete response artifacts, red pixels verified after those reads, or actual owner-replacement refusal. Those checks were never reached.

Tested repair

The successor fixture uses the small run-owned source-response.cjs handler: HTML is served as text/html; charset=utf-8 and modules as text/javascript; charset=utf-8. Source pins, wrappers, routing and the existing isolation headers are preserved. No observed hash was substituted for a pin. The successor also retains every actual source response before attestation, including refused responses, for diagnosis on a future attempt.

test-source-response.cjs executes the exact attempted server handler over loopback HTTP and reproduces the actual Chrome decoder mismatch. It then executes the repaired handler over loopback HTTP: all 41 complete responses carry the explicit charset, decode to their original pinned hashes and pass syntax checks. It also verifies Unicode HTML, isolation headers, missing-module 404, the successor's use of the tested handler and the corrected type2 call. source-response-test.json records the per-module results. The corrected 41-module preflight passes again. This is actual HTTP transport testing, not a mock source-response claim or a second Chrome attempt.

The original hardware gate remains unchanged. gateEligible:false means JavaScript ownership does not satisfy its four guest/render WASM contract. An explicitly named Software counterpart must wait for successful real browser ownership validation. It must preserve guest WASM identity, actual JS decoder/compositor owners and hashes/current epochs, drained parallel reads, final colocated owner/pixels, verified host clock and <=250ms age, changed timer pair and phase sinks, and sticky UNKNOWN without replay. No observer bypass, renderer WASM fabrication or contract weakening was introduced here.

Runtime cleanup and recovery

The supervisor prearmed an observable independent GNU timeout watchdog with an immutable 25-minute deadline, plus an inner 90-second diagnostic limit, before allowing Chrome to start. Only the actual runtime held /tmp/wine-runtime.lock. The browser and loopback server closed promptly on refusal. The subreaper recorded descendant/adopted PID/start identities and reaped adopted exits. Independent drain checks verify all 23 recorded identities absent, supervisor/watchdog absent, both loopback ports refusing connections, no process command line or open descriptor referencing the owned Chrome profile, and runtime lock free. No native input was dispatched. Process discovery is sampled at 30ms; the record is not an exhaustive history of short-lived descendants between samples. No stronger lifecycle claim is made from that sampling. The surviving-process/profile/port checks found no residual owned resources.

No new box exists: ID/state/expiry/stop receipt are all N/A. No remote box, original game, native input, public deployment, old-box resume/delete or subworker was used. The inherited archived box was untouched.

Recovery records exact checkout SHA in source-head.txt, all 851 existing providers, relevant Git diffs, small owned helper patches and existing media paths/hashes. No source-tree or media archive was created. Evidence is ignored; only this note is committed. artifact-manifest.json is independently reread and hashed before publication. Earlier failures remain retained.

Next: in a separately authorized bounded local attempt, run the repaired UTF-8 successor and obtain four complete uncached JS source responses over two current-owner reads, actual red pixels, common context/epoch and real owner-replacement refusal. Only then implement and test the explicit Software gate. Full27, parity, audio and FPS remain open.