Tunnel retained module and queued stream failure — 18 November 2026

No new original Tunnel or browser shutdown coverage was obtained. Original preparation refused the 1,520,000,000-byte reserve before starting a supervisor or Chrome. That original scope is closed without replay. The separate authored browser regression timed out during Chrome launch, before obtaining a DevTools endpoint or executing stream work; its scope is also closed. The startup cause is unknown. No moving frames, ordinary title close, native guest exit, browser crash recovery or browser CLOSE/STOP pass is claimed.

Evidence: scratch/runs/20261118-dx-tunnel-retained-module-browser-shutdown/, with result.json written last. The predecessor 20261118-d3dim-current-browser-shutdown-crash-sweep-1220 likewise never launched a browser; the earlier software Tunnel observations in dxsdk-software-tunnel-flip-rendering-finish-20261117.md remain historical evidence only.

The retained module at scratch/runs/20261118-static-display-fresh-capture-gltron-gameplay/tested.wasm is present, 2,418,710 bytes, SHA-256 a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8. All 1,453 original runtime/source pins, including 161 compiler/source pins, recover exactly from published Git f6962ba5b1cf2829a66f5cc4d80b293c30719b37. It was not recompiled or copied. This is a retained f696 module, not a verified current HEAD module: src/main.watx and src/14-cli-command-line.wat changed subsequently. No matching current HEAD module was established. The planned original graph used the same published renderer repair 06f27181f9c977cd1aed99270040efddca00ed75 as Quake II/GLTron, with other files from f696. Renderer/admission ownership and foreign WIP were left untouched.

A small run-local browser regression was implemented against unmodified production d3d-render-worker.js and GPU stream v2. It publishes native CLEAR/READBACK/PRESENT, an unknown opcode, and trailing readback/presentation, then immediately queues CLOSE. Assertions cover the exact failure sequence/error, completion stopping before the bad record, untouched trailing output and pitch canaries, pre-CLOSE frame ordering, a post-CLOSE transport barrier, same-ID replacement, queued STOP and a post-STOP barrier. Browser execution remains unverified because startup failed. This retained-input test belongs in the run folder; no tracked test reads run evidence.

The same regression body passed once under a Node worker_threads browser-API shim using the retained production WAT module. Actual stream failure was opcode 65535 at sequence 6/error 1, completion stayed at sequence 5, and only stamps 501 and 503 arrived; the trailing stamp 502/readback did not publish. Native pixel and pitch-canary assertions passed. This extends the relevant Node queued failure coverage without claiming browser behavior, an original game crash, or ExecuteBuffer lifetime coverage. All eight Node production dependency files match the selected published source; the shared foreign region mirror differs, but its producer heap base equals f696 and was recorded explicitly. The browser server selected the published mirror.

Both finite native-supervise command scopes have actual successful supervisor cleanup with no leaks. Browser preparation command exited 1 after the 15-second startup timeout; its supervisor wait records both pipe EOFs. Chrome launch failed before the harness could attach its own wait, so a separate Chrome wait/exit status is unavailable. All fourteen browser-scope tracked PIDs and the Node driver PID were independently absent afterward. The Node driver exited 0; its Worker was explicitly terminated by the harness and actually exited 1, distinct from guest exit. The owned tmpfs browser workspace was removed. The finite task disk/board monitor is joined separately before final sealing; its sampled floor outcome is recorded in monitor.json. Earlier transient disk-floor breaches remain unresolved; no storage cause or repair is asserted.

The final finite monitor actually joined with exit 0 after 3,915 samples. Its minimum was 1,463,742,464 available bytes, with 80 samples below the 1.5 GB floor: an actual task-lifetime floor breach, not a compliance pass. The separate authored browser watchdog did not fire and sampled a minimum of 1,521,725,440 bytes during its 16.7-second command. The transient global allocation cause remains unknown. This run's retained outputs total approximately 0.5 MB; no foreign cleanup was attempted.

No production code or tracked tests changed, so no canonical build or full focused suite was run. The run-dependent narrow test and syntax checks are the validation here; the existing ExecuteBuffer test requires a separately instrumented compile and was not rerun. Recovery is Git plus recorded relevant binary diff and owned script/report patches, with the unique module retained at its existing path. No source snapshots, media copies, screenshots, boxes, workers, Python, Wine or deployment were created. Box ID null, state NOT_CREATED, expiry null, stop receipt null. Release and stop.