Asghan: native attack input reaches guest polls; outgoing consequence unresolved

The unchanged registered software demo consumes Ctrl+Up during a reviewed contact-range attack window. Outgoing sword damage, enemy defeat and pickup remain unproved. This run closes the input-delivery question, not the combat consequence question. No demonstrated emulator mismatch warrants a repair.

Evidence: scratch/runs/20261108-asghan-native-attack-consumption/. Tested main SHA: 924096418d11b668ca0341b101133d32cb96370c. Recovered executing guest Worker module SHA-256: f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. Only this note changes in Git; run helpers and evidence remain ignored.

Controls and actual input consumption

Read CLAUDE.md, the current browser-action handoff, and shipped readme.txt sections 3–4. The documented controls are hold Ctrl for stance, Ctrl+Up for fast attack, Ctrl+Left for swipe and rotate, and Space for action/pickup. The run used title → Enter → visible NEW GAME → Enter → settled pier, then reviewed Up/Left movements toward the same visible orc and barrel.

The released native browser driver, independent release timer, exact ACKs, owner/window/canvas/module binding and 120-second scene freshness gate were retained. Before the original run, trusted native Alt, Alt+Up and Ctrl+Up DOWN/UP consumption passed against the actual production Calc Worker. Generated owner/window/module/image/stale/UNKNOWN negative controls passed.

There was no suitable existing bounded Worker-return recorder. A run-local CDP observer was installed before the production Worker instantiated the unchanged module. It forwards the existing WebAssembly.instantiate call and wraps its actual host input imports using Reflect.apply exactly once. It preserves receiver, argument values, returned value/promise and original exception identity; its recording failures cannot replace production results. Receiver/result/exception/exactly-once checks and actual production Worker native-input qualification passed (observer-qualification.json). It makes zero additional input polls, writes no guest memory and changes no clocks, scheduler, executable or setup. Observation has overhead; this is no speed test.

get_key_down_state returns, including Worker-local shared-bitmap/cache answers, were observed at the actual imported-function boundary. The sweep of keyboard keys plus di_keyboard_buffer calls is consistent with the DirectInput keyboard path in src/09a8-handlers-directx.wat; attribution to that named guest API is a source-based inference. The measured facts below are actual import results, not inferred browser dispatch or stance pixels.

Reviewed attack Native hold Guest returned down Guest returned released High returns since preceding capture
contact-fast, Ctrl+Up 04:45:17.337–19.337 UTC Both 0x8000 at 17.472, 135 ms after DOWN Both zero at 19.520, 183 ms after UP Ctrl 18, Up 8
contact-swipe, Ctrl+Left 04:45:36.516–38.516 UTC Both 0x8000 at 36.533, 17 ms after DOWN Both zero at 38.550, 34 ms after UP Ctrl 16, Left 7

The counters count import returns, not frames or distinct physical presses. Native DOWN/UP ACKs completed within 1–2 ms; independently armed releases were attempted at their scheduled deadlines. The full transition rows, counts and pulse receipts are in attack-consumption.json, recovered state files and recovered/original/final-consumption.json. The observer reports one instantiation, zero errors and zero dropped transition rows.

What the reviewed images establish

approach-orc.png shows the orc beside the barrel, ahead-left of Asghan. close-range.png shows that forward movement instead takes Asghan onto the raised edge beside the rock wall, with the orc lower and left. face-target.png follows a left turn: the orc's axe reaches Asghan and red ground marks accompany a shorter hero health bar. This establishes contact reach for the enemy; it does not establish the hero's sword alignment.

One two-second fast attack follows that current reviewed capture. In contact-fast-during-900.png the sword extends across the orc's head/upper body in projection; the orc subsequently remains standing and armed in contact-fast.png. Sword motion and blood are insufficient to attribute damage. One documented swipe-and-rotate then addresses the oblique contact angle; contact-swipe.png still shows a standing, armed orc. No disappearance, loot, inventory/counter change or clearly attributable hit reaction was observed. No visible pickup was selected, and Space was not sent blindly.

Displayed green pixels at screenshot row 666 measure 160 → 130 → 100 → 90 in face-target, contact-fast, contact-swipe and the final capture. These are hero-bar pixel widths, not private health values or enemy damage. There was no death/restart joined into the sequence. After these two distinct attack attempts, attacks stopped. The final capture is labelled native-menu after an Escape request, but no menu overlay is visible in that image; it is not menu-success evidence. Chrome then closed normally.

The remaining measured gap is target alignment/vertical contact on the raised edge and attribution of an outgoing consequence, despite Ctrl/direction being returned down together during the hold. Missing Ctrl/Up delivery does not explain these attempts. The evidence does not establish whether the sword missed, damage occurred without a visible reaction, or additional guest combat rules applied. Do not repeat these poses as success. A future attempt needs a reviewed same-plane target directly ahead, or a visible actionable pickup with a before/after object or inventory consequence.

Identity, preservation and shutdown

All 120 fixture files, 30,448,512 bytes, match the immediately preceding run's path/size/SHA-256 references. No install, native setup edit, alternate executable or media archive was used. All 712 runtime pins match before launch and after closure. HTTP responses independently match 69 originals, including _start.exe, both required DLLs, manifest and native setup outputs; 51 paths have no recorded HTTP request. Staging is not described as consumption of those 51.

Every one of 13 bound captures retains the same process 1000, actual guest endpoint/Worker/instance/module, page renderer, canvas and visible process-owned Win Alis window identity. The route remains software/DirectDraw presented through CanvasRenderingContext2D, with no owning 3D Worker/context. No GL-only readiness gate was used. This supplies no GL/D3D, FPS, audio or full-playability claim.

verify.cjs passes: 13 exact command ACKs, 10 pulses, 12 balanced trusted native DOWN/UP pairs, zero UNKNOWNs, 40 CRC-decoded PNGs, stable capture owners, actual attack-poll transitions and unchanged media. All 183 remote references, 14,017,036 bytes, were retrieved and independently SHA-256 checked before normal box stop. Chrome exit 0; controller/Chrome PIDs and both ports absent; profile/writer lock removed; no held keys. Source recovery records the Git SHA, relevant binary patches, owned-input patches, recovered module and original media references, without any source-tree archive.

Fresh UUID-idempotent box bx_8t63hy52 had an independently observable 25-minute watchdog armed before allocation. Normal stop receipt stop_251149051c2cca22 completed at 2026-11-08T04:46:38.400Z, forced=false; final state archived, expiry null, initial expiry 05:08:40.693Z. Both lifecycle delegates exited 0 and the supervisor has no remaining children. Disk stayed above 5 GB. No RAM gate, old-box resume/delete, nested worker or public deployment occurred.