Q3 arena harness and native runtime, 1748
The arena scheduling harness now includes the generated startup diagnostic
prelude together with send. The former function-only extraction left
startupOutbound and startupSettlement undefined and produced an empty wire;
that was a harness incompatibility, not evidence of a production transport
failure. No exception suppression, replacement send/receive implementation,
production timeout, routing, auto-attach or debugger-policy change was made.
Harness fix published on main as d829d3545. Evidence belongs to
scratch/runs/20261109-q3-arena-harness-runtime-1748/. Recover source using
gitbase.txt and relevant-uncommitted.patch; Git holds the source archive.
Only the four small run-specific Python inputs and generated authenticated
callers/runtimes are staged. Media remain referenced by path, SHA-256 and size.
Checks
- Seven actual composed command/screenshot scheduling cases pass: baseline losing continuation, fixed late screenshot, overdue response, overdue owner, success before deadline, first protocol error, and untimed fixture. Exact outbound command envelope hashes, session/method/request identity and one settlement per request are checked using the generated diagnostic functions. First failure survives later command refusal; no input or screenshot continuation is admitted after terminal failure. The scheduling harness retains inert owner replies; it is not a complete browser implementation.
- The separate full authenticated generated-relative-caller peer suite passes all 16 cases, including transformed transport byte identity, send return, public inbound, promise settlement, callback/error/throw identity, observer failure isolation, record/timer bounds, late unmatched replies and unchanged 20-second request failure. Those peer timings are not native measurements.
- Current foreign closure: production preparation, staged generated caller, Git-free actual consumer compilation and 27 refusal controls pass.
- Both generated native lifecycle prelaunch/custody/staged checks pass before Chrome; actual generated owning-caller integration and fresh interactive controller tests pass, including distinct wait/EOF and independent watchdog joins. These offline waits do not establish native browser cleanup.
flock /tmp/wine-build.lock bash tools/build.shpasses (build.log).
Native run
Fresh preparation independently SHA/size-verified 942 providers. Manifest
SHA-256 is 949e95016d5bf2a863716b835f7432e1002f02b22da5b4a30790572fed70c64a.
The actual CPU and renderer module identity is
3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18;
compat remains ecb45716b1ee7fc5218b1670dd270b19c7664ee30d000b3360a12786a1c21eaf.
This assignment explicitly authorized the runtime after Nox1733's no-box final release and ReturnFire's verified archived stop. Historical ReturnFire missing terminal/watchdog joins remain FAILED and were never inferred from absence.
One box, bx_7de5kh4r, was created at 17:50:59.313Z. Its unique watchdog UUID
is 17ced03f-eeb2-4328-989c-8859b48b0edd; the immutable deadline was armed and
independently observed live before allocation. Initial archiveAfter is
2026-11-09T18:15:58.170Z; expiresAt was not returned.
The single changed fixture passed production-worker/module ownership, nine negative controls, positive owning WebGL draw/presentation, physical held-key and release checks, and actual screenshot/ACK0. Planned owning cancellation passed separate generated-parent/prelaunch/runtime waits and stdout/stderr EOF; kernel custody reaped 164/164 identities. Physical inspection passed with ports closed, profile absent and locks drained. Normal browser close remains UNKNOWN for this deliberately cancelled fixture; kernel reaping is a separate fact.
The conditional original reached the main menu, CHOOSE LEVEL and DIFFICULTY.
Each action was reviewed against its actual captured image. An early read-only
download of arena-fight-capture.json returned 404 while the command was still
pending; after its successful ACK, the published image was retrieved and hash
verified. No browser command was replayed for that inspection mistake.
The original returned successful ACKs 0–8 and nine native captures. The
start-arena screenshot showed CONNECTING TO LOCALHOST; the next wait capture
was unchanged. arena-settle.png then shows the textured Q3DM1 courtyard,
weapon, bot and HUD (100 rounds, 119 health). This is verified current original
world rendering, not a fixture promoted to gameplay. After a balanced 1500 ms
W command, move-forward.png shows a death scene and “Fragged by Major.” That
camera change cannot establish useful forward movement. Firing is unverified.
The concrete terminal cause is command admission, not a screenshot timeout:
respawn-fire failed at 17:59:10.542Z with command cannot fit immutable
run/cleanup deadlines; not queued. The actual original browser stop was
18:00:29.024Z. The unchanged owning caller passes a 55-second minimum ACK
budget, and transport reserves 30 seconds for cleanup. Its latest admission
was therefore 17:59:04.024Z, before this request. No command 9 was published,
no respawn/fire input was dispatched, and no later input/capture/retry ran.
Owning transport cancellation followed; local controller exit 1 and every
separate wait/EOF are retained. No deadline or admission rule was relaxed.
Transport evidence and limits
transport-analysis.json independently checks compressed SHA/size, decompressed
SHA/size and gzip CRC, plus exact outbound/public inbound byte identities.
For page-session Target.setAutoAttach request 6, the changed fixture observed
159.363069 ms from synchronous send return to public inbound, and 159.727366 ms
to consumer resolution. The original observed 0.728128 ms to public inbound
and 1.521038 ms to resolution. Request/session identities match. The previous
startup stall did not recur in these two sessions; this does not repair or
explain the historical stall.
The bounded startup observer stopped at successful navigation, before the first 250 ms scheduling tick: both sessions have zero native Node scheduling samples. No claim about later Node stalls follows from that absence. Later input/screenshot diagnostics are retained separately; startup observations do not cover arena screenshot transport. Send return is API return, public inbound follows buffering/decoding/Node dispatch, and settlement is consumer promise settlement. None proves socket flush or browser-internal receipt/dispatch.
Recovery and shutdown
All 76 fixture and 143 original unique-per-session compressed artifacts were
retrieved and independently reread before delegated normal stop. The global
unique count and complete SHA/size/gzip-CRC audit are recorded in
transport-analysis.json. No source tree or game-media archive was made.
Original kernel custody reaped 355/355 identities; independent inspection
passed with both ports closed, profile absent, no runtime FD holders and locks
drained. Browser ordinary close/EOF remains UNKNOWN: owning cancellation and
kernel reaping are separate from normal browser completion. The local
interactive controller actually waited exit 1, separate stdout/stderr EOF,
unforced. Its launch transport's -9 is retained, not relabeled normal exit.
Box bx_7de5kh4r is ARCHIVED, unforced completed stop
stop_92a6058ee1659620, ended at 2026-11-09T17:59:54.128Z, with final
archiveAfter null and expiresAt not returned. My immutable UUID watchdog
observed that normal-stop receipt, wrote its terminal record and was actually
waited with exit 0 and separate stdout/stderr EOF. No box remains billing idle.
Next: preserve these current world/transport observations and plan original input admission from the actual browser deadline early enough to test movement and fire before the existing 55-second ACK plus 30-second cleanup boundary. This task ends with the tested published harness fix and a verified current world screenshot plus a concrete pre-dispatch deadline refusal; movement/fire remain unverified. No emulator causal repair is claimed.