Nox menu acceptance: missing consumer state prevents a causal repair

Run scratch/runs/20261109-nox-menu-acceptance-cause-2348/ stops offline. The first concrete blocker is the absence of the mouse consumer state needed to reproduce the demonstrated Conjurer stall. A route difference is established; its causal effect is not. No production driver change, allocation, native fixture, original launch, or new gameplay result is claimed.

Nine retained images were individually model-viewed and SHA-256 checked against their capture records. The successful 2255 and failed 2325 runs have byte-identical step-009-flag-place.png images (SHA-256 5f612fd198573648d508f90e3a5fa016a9d09b1ecc8ea580d26f6d4235d6745b). They show selected Conjurer, visible Accept, and the guest cursor near screen (620,470). Their first nine actual commands also match exactly. The successful next command was native click [0.025,0.477]; the failed run instead first clicked [0.5,0.794], then tried [0.025,0.477], then Enter. The first failed click leaves the guest cursor near (620,470); the retained point subsequently moves it near (465,550), above Accept. Enter preserves that selection screen. The successful image instead reaches default Jack customization. This is a stateful route divergence, not evidence that the retained point stopped working in an otherwise identical consumer state.

Both recorded canvases are intrinsic 640×480, with CSS rectangle 945×855 at (0,34). The generated mouse helper maps [0.025,0.477] to CSS (23.625,441.835) and the browser bridge's logical canvas (16,228); [0.5,0.794] maps to CSS (472.5,712.87) and logical (320,381). These are not guest software-cursor coordinates. The two retained-point mouse receipts record the same CSS destination and 150ms hold. Native mouse, binding, and key helpers are byte-identical between runs. Generated consumer source differs only in its immutable plan digest, as checked after normalizing that one digest. Both native capture identities bind the same primary module, SHA-256 3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18. No driver logic regression is demonstrated.

The production lib/browser-input.js bridge can suppress absolute movement while relative capture is pending, and uses relative counts plus the renderer's virtual cursor when locked. Task-local controls execute this actual bridge with the actual generated native mouse helper: the same centered click reaches logical (320,381) in the absolute case, produces no absolute move in the pending case, and clicks the virtual (420,300) in the locked case with explicitly zero movement counts. Those pointer states/counts are synthetic controls, not recovered facts about either native run. The retained evidence does not record pointer-lock state, the exclusive presentation transform, virtual cursor/DI reference state, DOM movement counts, or guest DirectInput consumption at the failing commands. The archived terminal run must not be resumed or queried to fill that gap.

The existing offline Node peer in tools/nox-current-driver.test.cjs directly emits listener events; it does not run the production canvas input bridge, renderer routing, or Nox's input polling. A passing peer route therefore cannot validate menu acceptance. The task-local route controls explicitly run both the 37 successful actual requests and all 12 failed actual requests through the current generated consumer, preserving holds and using zero post-action waits and the peer's command-N transport IDs. Their returned identity/review chains are checked. Neither result can establish guest navigation.

The locked build, three native-input tests, all 27 generated-consumer controls, and five additional whole-consumer binding controls passed. The latter alter only external review/image reads: missing review, wrong SHA, foreign identity, foreign label, and replaced PNG. Each asserts zero mouse presses, startup-only capture, no Browser.close request, no protocol after terminal publication, and separate child stdout/stderr EOFs. Eight exact builder CLI/function controls check terminal markers, initial image/hash/label binding, review text and label construction; the 37-command builder output matches actual successful requests. The real browser bridge's three pointer-state controls and existing relative input, clipping and lock-gate tests passed. Both complete route controls passed, demonstrating the peer's inability to distinguish this actual guest failure. These are offline Node/control-flow results; none proves native guest repair.

One initial integration invocation ran before generation had published its manifest and returned local ENOENT. integration.log retains it; no peer or Chromium started. The subsequent complete invocation used the sealed generated files. This was test sequencing, not an input retry or a diagnosis of Nox. Recovery uses the exact Git base in git-base.txt, the generated plan's Git commit, empty relevant uncommitted-source diff, generated provenance patches, and small run-inputs.patch. Original media and retained images are referenced by existing paths and hashes. No source-tree/media archive was created.

The separately reviewed accepted staff images still establish the explicit Force of Nature Staff of Recharge label, equipment with charge 3, and green particles with charge 2 while the northeast door remains intact. They do not establish target damage. This task performs no staff recovery or attack. The next necessary evidence is passive observation of the actual input boundary (DOM coordinates/counts and lock state, presentation/virtual cursor, DI delivery and consumption), bound to the reviewed Conjurer image, sufficient to select and causally test a changed route. Replaying either unchanged failed sequence or interpreting a transport ACK as guest acceptance would not resolve it.

Box ID, state, expiry, watchdog UUID and stop receipt: not applicable; this worker created no box. No owned live runtime remains.