Redline menu caller: CSS jumps became guest mouse counts — 2026-11-09

The original 0955 Training click consumed [-303,-231,0,128] as native mouse state, rather than selecting the photographed Training button. The run-specific click helper always sends absolute CSS positions; after Pointer Lock the real browser bridge forwards their differences as unscaled guest mouse counts and routes button events through the virtual cursor. This is a caller protocol defect. The correction stays in a Redline-only helper and small run-input caller patch; shared clients/host/WAT/custody remain untouched.

Exact retained caller/helper/browser/renderer/app source pins and six PNG hashes match. New Game requests CSS (481.95,326.41) at rect (0,34,945,855), backing 640×480. Trusted events show unlocked move/down, lock acquisition, then up. Its visible result is the mission chooser. Its native mouse buffer contains zero movement and button128: this success does not validate absolute cursor placement. Back subsequently requests (784.35,769.3) while locked, producing (303,443) counts; back-menu.png remains the chooser. Escape returns to Start Up, but the retained event state is still locked.

Training requests (481.95,538.45). At 09:58:14.758Z its trusted locked move is (-303,-231), with frozen DOM anchor (481,326). At .761Z the owning original mouse-state buffer is [-303,-231,0,128]; subsequent samples are zero motion with button128, then release0. Guest WM_LBUTTONDOWN/UP at .788/.909Z independently carry client (322,225) (lParam14745922); associated nonclient point is (325,248). Training-result and terminal menu-down PNGs show Start Up with the game software cursor near the top-left, outside every menu button. Win32 and private game cursor positions differ and must not be conflated. Exact private coordinates and original menu handler implementation are unavailable. Down's three wholly-held VK40 highs/release0 do not establish menu-selection handling.

tools/redline-menu-selection.cjs refuses absolute targets while locked. For an explicitly requested cursorOnly click it uses the caller's last submitted CDP point, preserving zero displacement, and reuses the established native helper's 150ms independent release. Missing lock/position or an invalid hold fails before effects. A down failure still attempts up and propagates. The run-input prepare-caller.cjs recovers exact Git caller922fae307 and adds only current lock/last-CDP-position arguments at its existing mouse callsite; native-mouse.cjs connects the new helper. Existing owner admission, freshness, identity, deadline, input bounds and first-failure terminal gates stay intact.

Supported ordinary selection recipe: with a fresh reviewed Start Up screen and pointer lock, use the existing relative action (at most24 counts per axis) to move the visible game software cursor onto Training Levels. Derive each step from its reviewed hotspot and game-image scale; review progress after each bounded step. Once the cursor/hover is visibly inside Training, submit {"action":"click","cursorOnly":true,"point":[0.5,0.5],"holdMs":150, "waitMs":1000} with the fresh review reference and a new result label. The point is a placeholder; the adapter clicks the last CDP position without moving. Review the actual Training list before targeting Vehicle Training. Unexpected cursor motion or any failed/unknown result is terminal. No absolute target jump, Down fallback, or resumed 0955 stream is permitted. This recipe has not been native-tested. The initially considered unlocked-only shortcut is explicitly rejected: zero native motion on successful New Game makes it insufficient.

The composed offline test executes the actual old/corrected caller mouse branch with real helpers: oldFAIL is exactly the unwanted coordinate jump; newPASS sends move/down/up all at the previous CDP point. Focused refusal/release tests and real browser-bridge/relative-input regressions pass. These establish protocol correction, not native Training/world/vehicle progress. The required locked build passes. Shared free disk briefly falls to1499844608B after build (below1.5GB); the deviation was reported and later observed restored above 1.5GB. Unique logs are gzip streams; evidence is independently SHA/size/CRC reread and patches reverse-check. Evidence, recipe JSON, source recovery and checks are in scratch/runs/20261109-redline-menu-input-cause-1010/. Before any separately authorized runtime, integrate/re-pin the full unchanged dependency closure plus new helper/caller inputs and run its required fixture; no launch is performed here.

No box, Chrome, guest, agent, worker or deployment was created: box/state/expiry/ stop = null/NOT_CREATED/null/null. Prior bx_n6m3yphy / UUID257e096f-24ad-4df8-9520-0d65f924d3c3 remains ARCHIVED/PHYSICAL_DRAINED; unforced stop_65b7f445cf93d6e8 completed10:01:10.317Z, initial expiry10:17:38.402Z. Original first actual failure remains terminal. Native lifetime REVOKED; all-writers/edict ABA UNKNOWN. SingleTaskSTOP.