Quake II guarded native run — 2026-11-14

Run: scratch/runs/20261114-quake2-guarded-enemy/.

The requested outgoing hit/defeat and onward traversal were not established. The original registered quake2_demo reached the lower opening-room floor and slatted passage at 100 HP. No living soldier or attack was reached. Native Customize Controls, Pause and F1 Help were personally reviewed. The last turn was positively cancelled without consumption at its five-second scene deadline. The frozen controller closed further input. Orderly host shutdown exited 0; native Quit, clean native Quit and Full27 remain unproved.

Controls, native review and progress

The original pak's default.cfg and bundled lib/quake2-modern-controls.ini were read without extracting or editing game media. The overlay changes Enter to attack and Ctrl to crouch. The live F4 Customize screen independently confirmed attack ENTER or MOUSE1, duck c or CTRL, WASD/arrow navigation, PGDN lookup, Z/DEL lookdown, END center view and F1 help. No binding was selected for editing. Ctrl was never sent.

Native Pause displayed PAUSED and held the opening scene for manual review. F1 requested while paused appeared after unpausing: Outer Base, easy, kills 0/18, goals 0/1, secrets 0/3. The help overlay was then closed.

Finite reviewed keyboard movement cleared the diagonal support, reached the lower floor and entered the slatted passage. Settled captures were used because some immediate post-pulse images preceded the eventual camera change. Several short right-turn handler acknowledgements did not produce a corresponding guest key-DOWN observation or visible turn; no generic emulator defect was isolated or patched. Navigation remained in the passage, without a reviewed next-room or enemy encounter. There were no attack pulses, unchanged Mouse1 bursts, outgoing-hit claims or weapon-pose-based combat claims.

result.json names exactly nine retained gameplay PNGs, including native bindings, PAUSED, F1 kills 0/18, opening room, passage entrance/interior and the terminal scene. Other manual captures were removed after recording their selection hashes. No automated frame series, source archive, media copy or WASM copy was made.

Guarded input and the concrete blocker

The new tools/quake2-controller.py applies an irreversible 180-second Quit reserve, exact reviewed Quit routes, native-binding-review admission, five-second aim scenes and thirty-second menu scenes. It clamps the actual production guard deadline to scene expiry as well as the gameplay cutoff. A focused negative fixture fails against its preceding policy; all eight final policy tests pass. This is a controller safeguard, not a demonstrated guest runtime repair.

At original launch, tools/quake2-guard.py was an independent derivative of LBreakout's tested shared helper, with only its title and documented Quake II key whitelist changed. The shared Forsaken controller, supervisor and tests were read-only. Their exact tested hashes and original derivative delta are preserved. The live runtime closure was not edited, rebound or relaunched.

Actual initialized native GL, a submitted full scene and personal image review preceded the first production guarded-bind. The permanent owner was tied to native PID 3807829, start token 272854922, and its exclusively owned loopback endpoint 8557. All gameplay edges used production guarded input, unique nonces, one-use captures and independently prearmed UP processes.

There were 27 manual intents: 18 handler-consumed, eight positively refused before DOWN, and one final CANCELLED, consumed=false receipt. Some refusals followed tool round trips longer than five seconds; one occurred after UP had been prearmed but before DOWN. Fresh scenes and new intents followed only the positive pre-DOWN refusals. No nonce or capture was replayed. All consumed aim deliveries were within 4.517 seconds of capture; the menu maximum was 14.426 seconds. Maximum recorded completion after consumed DOWN was 1427.134 ms, below the 2000 ms bound.

The final far-end-turn receipt reports immutable deadline expiry, no queued effect and no consumption. Independent release was acknowledged and its child was actually reaped with exit 0 and both EOFs. This was not UNKNOWN. However, the helper's existing failed-pulse path writes terminal-input-failure.json even for this positive cancellation and disallows further input. That record was preserved; no clearing, owner replacement or guard bypass was used.

Handler acknowledgement is separate from actual guest delivery. The passive observer records 16 guest key-DOWN and 18 key-UP messages, versus 18 handler-consumed intents. Enter DOWN occurred only in the two ordinary menu selections. Visible native movement, Pause, Customize and Help establish those specific consequences; they do not establish combat.

After the original was closed and its source pins checked, an HTTP fixture reproduced the same unnecessary terminal closure. The owned derivative now accepts only explicit deadline-expired cancellations with matching owner, epoch and nonce, no queued effect, consumed=false, null consumption time and noReplay=true. Independent UP must succeed and a subsequent status query must match the cancellation. It stops the remaining edges of that intent and requires a new scene, review and distinct intent. Nonce tombstones remain. It does not clear any terminal record or relax the five-second bound.

Eight focused refusal fixtures pass: a positive cancellation followed by a fresh distinct intent; refusal of the first edge without sending the second; and terminal closure for a partly consumed pair, UNKNOWN, wrong owner, epoch, nonce or subsequent UNKNOWN status. Duplicate nonce rejection is checked. The first green attempt exposed a fixture capture/action name collision; that test naming error and its corrected run are retained. The earlier test loader invocation error is also retained, separately from the causal red. The post-run fix was never used for original gameplay or native Quit.

The next gameplay step requires a newly authorized original attempt to reach and review an actual enemy consequence and native Quit. This task did not launch a second original or retroactively qualify its missing outcomes.

Exact execution, build and recovery

Heavy checks and runtime waited for the reserved LM8 run and LBreakout persistence run's actual outer waits, then used /tmp/wine-heavy-runtime.lock. The full build passed under /tmp/wine-build.lock. Final checks passed:

After the refusal fix, the eight policy, eight refusal, eleven HTTP/guard and ten actual CLI/lifecycle checks, four production script suites and full locked build passed again. Their separate waits and final owned-file hashes are in postfix-*-wait.json and postfix-pins.json. These checks use fixtures, not another original Quake II process.

The unchanged registered startup was +set vid_ref gl +menu_main, with authentic DLL preloads, --headless-gl --real-ticks, no guest threads, default 1000-block batches and fresh VFS. No console/map command, cheat, solver, autoaim/tracking, guest code/state/config/clock edit, external save, download, root dependency change, box or deployment was used.

Build and launch HEAD: 69122540a296dbc90827f94c27e038d8663cd43d, with the recorded dirty/untracked Git patches. Actual loaded WASM SHA-256: c94ccf1b179f18bd4b47867445afed8cef2f1393b19aabfd259c7f3b94a7be09. Postflight found zero drift in 1690 source, 68 registered media/control, 162 provider and 104 observed loaded-file pins; the module also matched. Git plus runtime.patch, new-runtime.patch, original/launch-source.patch and operational input patches supplies original source recovery. Separate final-change patches preserve the post-run fix.

Provider recovery used the prior Q3 recipe's pinned package/lock locally, npm ci --offline --ignore-scripts, then exactly nine cached native files from scratch/native-provider-cache/node3d-20261114. All 162 provider size/SHA pins matched. Package SHA-256 was 3ed55cf48ef3a1a00e7971ea528b9b2c991decbbf8917901310d59f3808e936f; lock SHA-256 was 125ea670ef6333b93d4375ed558fc42c24af2da9a4fbe2ed49e929ee519e0976. An initial receipt-writing script variable error and corrected offline rerun are recorded; neither launched a game. The private installed dependencies were removed only after postflight; the shared nine-file cache was preserved.

Deadline, shutdown and actual waits

The single original launched 18:35:39Z, with immutable gameplay cutoff 18:52:39Z, runtime deadline 18:55:39Z and an independent 18:55:59Z watchdog armed before native launch. No deadline changed and no relaunch occurred. The final scene was personally reviewed, an irreversible Quit phase was recorded, and the closed controller admitted no native Quit key. The explicit host-control shutdown must not be relabelled native Quit.

The harness reports 699131 batches in 665.509 seconds, 3473 presents and gl_sw_enabled=0. Actual native wait was 0, with both EOFs and no drain errors, at 18:46:46Z. Watchdog, supervisor child, outer child and additional outer wait were all actual 0, with separate stdout/stderr EOFs; the watchdog did not fire. Native termination was about 533 seconds before the fixed deadline. All recorded process IDs were absent, endpoint 8557 was closed and the heavy lock was released.

The short launcher has actual child wait 0 with both EOFs; its invoking tool returned 0, whose separate stream EOFs are unknown. The additional detached waiter's own wait is unknown, although its actual child wait is recorded and its PID is absent. The first queue-only preparer was deliberately terminated with tool exit 143. The second preparer's tool also returned 143, reason unknown, despite all separately recorded checks returning 0 and its ready record being written. These tool outcomes remain separate from native waits.

No box was created: ID null, state NOT_CREATED, expiry null, stop receipt N/A. /home remained above 1.5 GB free. The task ends with outgoing enemy hit/defeat, onward traversal, native clean Quit and Full27 explicitly missing.