Q3 original menu input: verified route and evidence blocker

Evidence: scratch/runs/20261110-q3-native-menu-consumption-0135/. Starting Git: b56fdfc8ea2b8fb4c4da4dc48b71ca805acda483. Prior native result: durable release service.

Arena movement/fire remains unattained. This investigation found no demonstrated causal repair. No production code, input timing, completion condition, guest state, clock, shared provider or startup service was changed. In particular, published 7607cab84 and tools/q3-startup-custody.py remain untouched. No fresh box or native attempt was made: the task conditions another native attempt on an offline demonstrated causal repair, which is absent.

The read-only audit.py in the run directory independently decompresses and checks retained console and state artifact SHA-256 and size against the original retrieval manifest. It reads the existing EXE and PK3 in place, verifies their hashes, decodes the original vm/ui.qvm in memory, and asserts the relevant instruction sequences. audit.json records its results, exact input references, hashes, selected disassembly and the explicit unknown runtime boundary. No game media, source archive or recovered artifact copies were created.

The original EXE SHA-256 is 2c0dd4d06e3282e0abc41ebc8e84e2d9868a17182ac175b232a9bb604a8269a3; PK3 SHA-256 is e77abad2466f45a0a7ea018445528f9b95a0fe7789fa1abc1a7718bbf0754b08; UI QVM SHA-256 is 6903548472c2d95898c71c979304e07fbe966e9db2d22db5d9c1ee4f8ddeb88f.

What the retained original actually observed

Within the same page observer clock, command begin was 42154.955ms. Trusted Enter reached the production renderer down/up handlers at 42162.985ms and 42309.485ms. Host dequeues followed at 42352.315, 42354.145 and 42354.825ms: WM_KEYDOWN(13,0x001c0001), WM_CHAR(13,0x001c0001) and WM_KEYUP(13,0xc01c0001). The result image bound at 43950.875ms. These times describe page observations, not guest handler completion.

All nineteen owner fields match between startup and result. Execution slices differ, as permitted; focus remains HWND 65543. The prior native input receipt records balanced acknowledgements and two wholly-held physical samples. Neither physical samples nor removal from the host FIFO proves Worker or guest consumption. The installed observer explicitly records guestHooks:false. The first subsequent model-reviewed image remained mainmenu and caused owning cancellation. This report retains that terminal result and does not conclude that the guest ignored Enter.

Actual original handler path, statically verified

The game window procedure begins at 0x443a10. The separate procedure at 0x4433d0 belongs to the console window; it is not the game keyboard handler. The game procedure's ordinary keydown path calls MapKey at 0x443f80 and queues a down event through 0x4403b0. MapKey extracts the scan-code byte from lParam; the original table maps scan code 0x1c to Q3 key 13. WM_CHAR takes a separate queued character path.

The source route after the observed dequeue is Worker RPC input receipt and its scoped hwnd/lParam reads (lib/guest-rpc.js), the Win32 message APIs (src/09a5-handlers-window.wat), DispatchMessage into the actual window procedure, the engine event queue, and Key_Event at 0x40aa70. These are source/static routes, not retained execution observations. Key_Event tests the UI catcher before ordinary key bindings and calls VM_Call(3,key) at 0x40ac65. Thus the configured bind ENTER "+button2" does not by itself explain failure while the UI catcher is active. Character events are marked with 0x400 before the UI call and are not interchangeable with key 13.

Within the authentic UI QVM:

These bytes establish the intended handler route and disprove a static claim that ordinary Enter has no main-menu activation branch. They do not establish the active runtime menu pointer, cursor/item, flags, callback, UI catcher, VM execution mode, or which of these branches actually executed. The saved vm_ui "0" setting alone is insufficient to establish the loaded VM mode.

Validation and stopping condition

The read-only audit passes. Existing tools/q3-input-observer.test.cjs passes against the actual current generated consumer, including all nineteen owner field refusals, changing slice telemetry, stale image binding, malformed review and inactive observer refusal before dispatch, with production Enter down/up and host dequeues. Existing tools/q3-arena-command-boundary.test.cjs passes all seven actual composed command/screenshot clock controls through the current foreign closure: historical late continuation is reproduced, while the current consumer aborts without resend or late input. These validate existing guards; they are not failing-before/passing-after evidence for a new menu repair. The required locked build also passes.

Concrete blocker: the retained original contains no observation of Worker input return, Win32 dispatch, Key_Event entry/return, or UI VM key-handler entry/return. It therefore cannot distinguish an unhandled key from a handler still in progress or a completed handler whose result was not presented. Adding a wait, choosing a completion predicate, or changing delivery without establishing that boundary would claim a cause unsupported by this evidence. No new incomplete/ wrong-binding/first-failure guest-completion controls or causal native acceptance are claimed. The next investigation needs a narrowly bounded passive trace of that missing boundary, tied to the reviewed image, command and immutable owner; it must not infer completion from ACK, physical state, dequeue, or elapsed time.

Lifecycle: no task box was created, so no task UUID, expiry, stop request or watchdog was applicable. Existing stopped bx_hsh4576p and all other agents' boxes were untouched. No local Chrome, nested worker or guest execution was used. This single-task worker releases its files and stops with this blocker.