Aranna pre-gesture hover observer, 2026-11-06

The run-local observer now proves a native point and loaded target before the first target action in generated full-shell controls, for both shipped module variants. The subsequent guarded click visibly activates generated target A. This repairs the observation gap in b9e92d06c's menu-phase investigation; it establishes no emulator mismatch or original gameplay. Production is untouched. Evidence: scratch/runs/20261106-aranna-pre-gesture/ (ignored).

The actual movement handler 006ea740 calls the loaded control's vtable slot +5c at 006ea8ee and 006ea9b1, returning at 006ea8f1 and 006ea9b4. Immutable disassembly shows manager ESI supplying +78/+7c as the native stack point, with ECX loaded from the current list node +10. Neither return PC was in the previous observer's six click sites, and its top-kind1 requirement excluded ordinary hover. The generated predecessor invoked the click wrapper on DOWN. Missing hover packets therefore did not establish an emulator fault.

The new passive indirect-call hook leaves that click path intact. It admits only those two hover return sites after one-way arming, with ESI equal to the manager loaded through the existing singleton, the current frame's node +10 equal to ECX, the actual loaded slot/target, matching stack/manager coordinates and the existing target rectangle [280,132,517,178]. Repeated unchanged manager/object/point samples are suppressed. Kind24 still shares 32 admissions with coordinate stores (stores limited to the first eight), and the global ceiling remains 8192. Existing actual return-PC/SP transaction ends remain active. Only observer globals and bounded nonfaulting copies are added; there are no guest calls, state writes, input setters, execution switches or guard relaxations.

Standard observer SHA-256 is 2290e462da6e8220766a386c60274ecfd414742dc70c1df8f712fe5ce7736915; compat is 9fa86dbd23d2c41794b5ca0b05e27cc7fc6dc5141063cfd9debad7763e12512f. Same-closure disabled modules exactly match shipped bytes. Import contracts are unchanged; the predecessor's observer-only arm export is retained. Source identity, binary Git diffs, new-file patches and module source patches are retained; no source/media archive was made.

The normally closed standard2 and first compat owning-Worker full-shell controls prove (398,155) at generated receiver 00405a00, manager 00405900, node 00405e90, slot 00405c5c, loaded target 006f34f0. The arm marker is serial2816; hover transaction152 starts2821, records native arguments2822, target2824, manager/node2825 and balanced end2826 at 006ea8f1. Only two shared admissions are spent. Before the target click, guest hit=0, down=up=1 (the earlier center acquisition outside the target). Only after this retained proof does command3 click, yielding hit=1, down=up=2 and personally inspected ACTIVATED TARGET A text. Target/session/module identities are pinned in each proof. The full-shell pre-action coverage is 006ea8f1; the other return site is exercised in the expanded passivity matrix.

native-control-proof.json retains 20 rejecting mutations per variant: owner components/module, stale phase/movement, wrong native or manager point, wrong rectangle/loaded receiver, missing target/provenance/end, wrong return PC/SP, reordered end, action before proof and saturated shared/global quotas. Both captures have312 completed transactions and zero incomplete ones. Two eight-cell matrices compare disabled/observed × standard/compat × ordinary API/owning Worker: CPU/lazy flags/x87 payload/environment, API/COM/SEH, guest results, immediate/buffered mouse and balanced native input all match. The expanded matrix actually observes both hover sites in each observed cell, reaches the unchanged combined32 ceiling and stays below global6707. Results are losslessly gzip-compressed with verified decompressed hashes and restoration recipes.

The first generated attempt rejected pixel arguments to a fraction-valued canvasclick before input. A later compat command correctly refused a standard-default module selection; the command writer now selects the requested mode. The first expanded matrix attempt rejected an existing HTTP receipt filename before browser launch; a distinct output name preserves both attempts. Two standard capture-review descriptions were inaccurate; explicit adjacent correction files preserve them. Compatibility previews also disagreed: activation text was visible, then a preview appeared black, including to the coordinator. All corrections remain. Two independent PNG decoders (stdlib and pngjs) establish identical activation-region bytes across standard2 response03, compat response03, compat2 response03 and fresh read-only compat2 response04: 1150 bright pixels, region SHA-256 51384adb9849a5c11a9e7307f11c186d1d8f389ea10d08b9bf43e3582d74ed93. The coordinator subsequently personally saw activation in compat2 and accepted both controls. The cause of inconsistent previews is unknown; retained PNG mutation is not established. Compat2 independently proves (399,155) before action. Its runner exec ended143 without driver/Chrome exit receipts; all three recorded processes were independently absent and its orphan profile removed. This uncertainty is not a normal-exit claim; final qualification uses the normally closed first compat run. None of these is an original guest failure.

Execution began only after the coordinator's21:24:11.941770–22:09:11.941770 UTC grant, predecessor FINAL and independent cleanup; queued preparation was read-only/tiny. No boxes were authorized or launched. The coordinator separately granted at most one original at21:45:47 UTC, after accepting generated controls; the one-attempt guard and exact grant are retained. Production focused tests/locked build are not claimed: this is a qualified run-local observer, with exact module compilation, passivity and full-shell tests above.

One separately authorized original

The exact21:45:47 UTC separate grant admitted one materially changed original. It accepted EULA, naturally rendered intros and reached Main Menu. An ordinary center acquisition preceded arm serial2675. A one-count movement measured native (398,399) before the next -244 Y movement, rather than assuming a previous run's position. The response12 screenshot visibly highlights Single Player. Its original pre-action-proof.json passes unchanged strict checks: point (398,155), receiver 0a4f40b0, manager 09e2e610, node 0a16fd9c, loaded target 006f34f0, slot 0073099c, transaction212 ordered serial2688/2689/2691/2692/2693, balanced return at 006ea8f1, five shared admissions. All212 transactions complete; none incomplete. These addresses are runtime evidence, not a semantic target-name assertion.

Command13 was queued after review and passed target proof/owner checks, but no target mouse-down occurred. The native click helper refused actual canvas pointer lock required. Response12 had lock true; retained lifecycle shows recenter expiry, releaseExpiredAutomaticLock invoking exitPointerLock at21:54:02.842 UTC, then actual null lock at02.848. Command guard receipt is15.529. This is an observed existing host mode transition, not a demonstrated emulator mismatch. The assertion precedes page.mouse.down; zero post-arm mouse-down/up events were observed. Response13 does not exist. Driver exited1, Chrome exited0, enclosing runner exited0. No Single Player transition, setup or gameplay was verified. No restart was attempted.

The response11 review initially misstated native X399; its adjacent correction and raw store evidence establish X398. The Y399 measurement and subsequent -244 movement were correct. A board line briefly included a speculative receiver suffix; an immediate exact-value correction preserves the authoritative proof.

Precise next-run prerequisites and recipe

Do not replay this original unchanged. First extend the generated full-shell controls to cover the automatic relative-to-absolute lock expiry between reviewed proof and action. Preserve current proof/owner/order/quota checks and add a final live mode precondition before dispatch, so a stale locked action is refused without terminating a recoverable command queue. Qualify both module variants and meaningful passivity/negative controls. Do not force/reacquire a lock or alter guest state to bypass expiry.

For an unlocked menu, qualify ordinary move:X,Y followed by a separately reviewed ordinary click at the same native point; determine coordinates through the existing presentation mapping and native stores, rather than canvas fractions or a previous run. The action adapter must authenticate click verbs as fully as nativeclick and ensure no intervening move or mode change invalidates the current proof. This is a required generated qualification, not authorization to run the original or an already-qualified click route.

After those controls pass, obtain a new explicit coordinator original grant and use a fresh one-attempt guard, exact source/helper/media/module identities, disk/locks checks and finite cleanup deadline. Load naturally, review Main Menu, wait for the actual stable input mode, arm with zero admissions, and use only that qualified ordinary movement route. Require a completed current-point/loaded-target hover proof and current mode before target action. Review the distinct Single Player response before ordinary setup navigation. Gameplay requires retained movement plus a separate distinct gameplay response; menus/setup are insufficient. Stop once on failure; never force transition, patch guest state, relax guards or infer an emulator fault from missing evidence.

Cleanup/terminal and publication identities are recorded in the run's final assessment and result.json (published last). No original gameplay claim is made.