Quake II tmpfs attempt, 2026-11-18

Run: scratch/runs/20261118-quake2-tmpfs-world-enemy-consequence/. Living-enemy damage/defeat and ordinary Quit remain unproved. The one fresh original closed at 11:41:53.213 UTC before any guest DOWN. Only the reviewed host fullscreen consent was delivered, with independent matching UP and physical zero. No endpoint request, replay or second original followed closure. The earlier accepted Game → Easy and four visible movements belong to the previous run; its existing corpse remains uncredited.

Old scope audit

The previous sealed run was examined offline only. All fourteen world gesture commands explicitly declared mode: play; the operator selected 5,000 ms. The four delivered world gestures had frame/request ages of 4353/4475, 4678/4765, 4403/4590 and 3739/3884 ms. The rejected world operations' text incorrectly said “fixed30000ms.” That was a diagnostic defect, not evidence that world inputs received menu deadlines. This run corrects the diagnostic and requires explicit play/menu scope. No world scene was relabelled menu.

Unchanged runtime and changed host storage

The served runtime is Git f6962ba5b1cf2829a66f5cc4d80b293c30719b37, with module SHA-256 a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8. All 67 registered media/control inputs match before and after. Registry, modern controls, GL arguments and the registered 10,000-block cooperative schedule are unchanged. Optional independent child Debugger recovery was omitted; executing-child module bytes/roles and physical GPU remain unknown.

The exact previous independent prepared release protocol was reused: ARM acknowledgement before DOWN, 1,750 ms fallback and maximum 1,500 ms hold. Disposable Chrome profile, cache, crashes and temporary files moved to a private /dev/shm/quake2-browser-* directory. Raw XWD stayed in memory; selected complete PNGs and metadata stayed on disk. Normal Puppeteer defaults and explicit component-updater/background-networking disabling were retained. The worker coordinated on the board and did not edit the tmpfs worker's helper or canonical launchers. No pause-aware design, polling change, clock change, forced focus, state/config/save/media edit or gameplay automation was added.

One native GL fixture checked the changed host path: personally reviewed blue → red after trusted Enter, independent physical-zero releases, ordinary Escape → native exit 0, actual child/Chrome/enclosing waits, and tmpfs removal. The existing fixed 600-second/two-scope authored aggregate was preserved; only one authored scope was used. Before launch the original retained its 180-second startup, 1,200-second total, irreversible 180-second Quit reserve and 20-second cleanup limits. No timeout was enlarged.

Precise closure cause

The first post-consent current-display request returned:

The unchanged verifyReadback conservatively rejected this combination with “ordinary publication completion, retained pixels or unheld display readback.” The shared disk/admission record closed permanently on that assertion before any guest input. The failure's closed: false describes the transport flag at that instant; it does not override the durable shared admission closure.

Source inspection and production renderer tests identify the contract gap: serviceCurrentDisplayCapture can finish an already displayed frame after the writer releases, without another publication. Metadata retains writer state at request, with no explicit clean-service completion receipt. The operator therefore cannot distinguish that safe completion from the combinations its extra metadata fence refuses. The production validator checks the owner/output and unheld state around readback, but this run does not retroactively admit the rejected frame or weaken that fence. A future repair needs separately qualified completion evidence or pre-request readiness; it was not validated on this original. This is a capture admission limitation, not proof of broken combat or a game crash. There is no claimed local runtime fix or completed enemy/Quit result.

Checks, resource limits and cleanup

The locked build passed. The corrected controller protocol suite passed, including replacement, UNKNOWN, independent release and freshness controls; all 14 production capture/admission tests passed. The first controller test invocation omitted required QUAKE2_TEST_OUT and failed before execution; the initial progress update claiming it passed was corrected. Run-local scope tests exercise the actual gesture body and corrected diagnostics.

The original driver and enclosing launcher exited 1, Chrome exited 0, native supervisor exited 1 and watchdog exited 0. Actual waits and both pipe EOFs are retained. The offline audit verifies 77 recorded runtime PIDs absent, 28 actual waits with both EOFs, nine PNGs decoded and four matching independent physical-zero UP receipts. Supervision reports no leaks; private displays and both tmpfs profiles were removed after their browser waits. These are host cleanup, not ordinary Quake II Quit.

An independent finite disk guard overlapped the original and offline drain. It measured 1,081,348,096 bytes free on /home after original closure, despite tmpfs browser state. Thus the required 1.5 GB floor was not maintained through drains. The responsible concurrent writer is unknown; no foreign cleanup was attempted. A second finite sampler overlaps final offline sealing. Neither sampling nor later recovery erases that breach.

Evidence retains exact Git pins, relevant binary diffs, run-local operator patch/seals, input hashes, selected screenshots, closure metadata and actual waits. Git remains the source archive; no source snapshot or media copy was made. result.json is published last after audits and final receipts.

Box ID null; state NOT_CREATED; expiry null; stop receipt null. No agents, workers, boxes, Python, Wine reference run or deployment were created. Foreign WIP, held game work and UT media/partials were preserved. Ownership is released and this single-task worker stops with the original permanently closed.