Chrome startup cause, real browser queued shutdown, and Tunnel

The Chrome startup blocker is identified and repaired in tools/browser-transients.cjs. The installed Chrome ELF launches a nested headless BetterChromium timezone lookup against ipwho.is under /usr/bin/timeout 15 before starting the requested browser. A bounded trace records that child starting at 12:39:01 and the requested profile first opening at 12:39:17. The original 15-second Puppeteer timeout expires during that lookup. Empty early stderr was real; the timeout's reference to stdout does not establish a pipe mismatch. The supplied switches were eventually honored, so their position after about:blank was not the cause.

Disposable browser arguments now include --bw-timezone=Etc/UTC, matching the test guest's ordinary UTC locale and the project's existing explicit timezone recipe. No launch/operation deadline increased, endpoint was synthesized, security setting changed, or warmup page executed. Shared common transport, renderer/admission, foreign WIP and media remain untouched.

Evidence is scratch/runs/20261118-d3dim-browser-startup-cause-shutdown-original/, with result.json written last. The original failed scopes from the predecessor were not reopened. This task's authored diagnostic, authored regression and sole original have separate finite supervision. Browser children are registered with the workspace wait immediately at spawn, before awaiting launch(). The failing diagnostic has an actual individual Chrome SIGKILL wait and both pipe EOFs even though no browser object was returned. The trace's inherited pipes needed the independent supervisor deadline; its individual Chrome wait is unavailable, recorded separately.

The narrow regression passed in a real Chrome 151 browser using the production d3d-render-worker.js, production GPU stream decoder/software executor, and retained WAT module. It publishes native CLEAR/READBACK/PRESENT, malformed opcode 65535, and trailing work before queuing CLOSE. Error sequence 6/code 1 and completion sequence 5 match; trailing readback and frame 502 remain untouched. Exact RGBA output and readback pitch canaries pass. CLOSE acknowledgement, post-CLOSE transport barrier, same-ID replacement, queued STOP and post-STOP barrier pass; only frame stamps 501 and 503 publish. Actual Worker target destruction is recorded after harness termination. No fake events or Node shim provide this result. Chrome exits 0 with an actual wait/both EOFs, and its supervisor reports no leaks. This is malformed-stream failure recovery, not an arbitrary OS/browser crash.

One original dx_tunnel ran after the shared disk recovered above 1.5 GB, using software rendering. I personally viewed complete 1280×1024 private-display captures focused-early.png and focused-later.png: gray textured receding rings and yellow/black radiation signs change orientation and position. The entire ordinary 320×320 application window and surrounding browser are visible. The images differ at 84,505 whole-image pixels over 1,660 ms; this is motion evidence, not FPS or parity. Initial captures show a black client despite real executor progress. A live inspection attachment changed the viewport from 1260×900 to 800×600, then an ordinary browser/guest focus click preceded visible output. Those two effects were not isolated; no renderer cause or repair is claimed. Initial black evidence remains retained.

The ordinary title close used trusted mouse down at page coordinates (327,32), with an independently executed mouse up. The reviewed closed.png shows the window absent. Durable after-close.json records running:false, exit code 0, process exited/0, no windows, both guest/render retirement flags true, no retirement error, and a real render-stopped acknowledgement followed by manager retirement. Native ExitProcess is suggested by code 0 already present when host stop() begins, but its direct guest/native exit log was not retained. Do not equate process-object bookkeeping or host shutdown with native exit proof.

The original enclosing harness then failed before final log flush, individual Chrome wait, display waits and final module HTTP receipts were saved. Its supervisor records direct driver exit 1 and bounded sweep of the owned Xvfb/openbox, with no leaks. The outer tool returned 143; its own JSON wait/pipe EOF receipt is unavailable. The independent monitor similarly completed with actual supervisor-recorded child exit 0, but its invoker returned 143 without the outer JSON wait. These are incomplete lifecycle evidence, not successful waits. All tracked PIDs are independently absent; the private display socket is absent, unique download/crash directories were empty, retained display metadata was recovered, and owned tmpfs was removed. The original scope is CLOSED UNKNOWN after the durable close observation: no replay or later endpoint requests.

Both runtime selections reuse published Git f6962ba5b1cf2829a66f5cc4d80b293c30719b37 and the existing 2,418,710-byte module SHA-256 a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8. All 1,453 retained source/runtime receipt pins match that Git graph. The original renderer is the already published 06f27181f9c977cd1aed99270040efddca00ed75 version. Served authored response hashes and observer suffixes are retained. No runtime module was compiled or copied. The mandatory locked canonical build passed for the helper code change; its current-tree output was not used as the original's module or substituted for a600 provenance.

Focused failed-startup/writer-drain, unique-download retention, and whole-XWD pixel/PNG conversion tests all pass (3/3), as does flock /tmp/wine-build.lock bash tools/build.sh. The retained-input browser test stays in this run; no tracked test reads scratch evidence. Recovery uses Git plus relevant binary diffs, owned new-file patches and the existing retained module/media paths, with no source snapshots or media copies.

Disk compliance did fail during this task: the finite monitor recorded minimum 1,495,732,224 bytes and 499 below-floor samples before external recovery. The authored regression's monitor also records a floor breach. The original was withheld until recovery; no extra reserve or RAM gate was added. Coordinator recovery and the separate attribution sampler remain separate evidence, not a cause repair by this worker. No boxes, workers, Python, Wine or deployment. Box ID null, state NOT_CREATED, expiry null, stop receipt null.