Aranna native relative control, 2026-11-06
Run: scratch/runs/20261106-aranna-relative-control/ (ignored). Source and helper versions, relevant binary diffs/new-file patches, media references, actual instantiated modules, browser identities, captures and failures are retained there. No media/source archive was made. The coordinator granted sole local runtime through18:18:47UTC including cleanup, after AoW1610 and WinHelp1635 FINAL. Preparation ran no tests/build/browser and changed only new run-local scripts.
Reproduced input mismatch and repair
An ordinary generated Win32 PE runs through the actual registered full shell, creates an800x600 exclusive DirectDraw mode, polls GetCursorPos and repeatedly recenters with SetCursorPos(400,300). It accumulates sampled displacement into its own UI point, with distinct A/B hit rectangles and rendered activation messages. The driver uses trusted native page.mouse movement/down/up. It never writes guest coordinates, calls renderer input setters, synthesizes DOM events, or invokes a game function.
At1001x800, the actual presentation image is661x496 inside661x568 with Y offset36. Trusted movement(-60,+96) entered handleRelativeMouseMove unchanged with guestCounts:true, but the virtual host point became(340,395), not(340,396). The guest accumulated Y+95 and activated B at(368,250). Exact movement qualification failed despite the correct target activation. regression-before.log reproduces the same395-versus396 failure without a browser.
The production fix in lib/renderer-input.js preserves original guest counts for DirectInput arithmetic and uses the native pixel center for the relative inverse transform. That prevents the final floor from losing a pixel at a fractional boundary. Native DOM routing, absolute pointer mapping, guest warp semantics, callback retention and delivery are unchanged. Only the reviewed host-file census entry changes alongside this source and its regression.
Focused relative-input, ClipCursor-edge, pointer-lock acquisition/expiry and latch tests pass. The locked full build passes; its initial census rejection and subsequent scoped review are preserved. The same native full-shell failing sequence now passes: UI(398,300) ->(428,155) after(+30,-145), A; then ->(368,251) after(-60,+96), B. DOM, host and sampled guest movement, balanced buttons, actual slot0/endpoint/target/session/module and target activation are asserted. The separate1200x700 horizontally letterboxed control also passed A/B. Generated GDI repaint can yield black/partial captures; each reviewed capture's actual contents and separately read guest state are distinguished. Review-prose corrections are retained explicitly.
Observer and semantics
The retained standard/compat disabled/observed ordinaryAPI/owningWorker matrix passes, including CPU/lazy flags/FPU/environment/fault/SEH and input equivalence plus missing/wrong endpoint negatives. The writer retarget adds passive store packets92/93 and shares the existing kind24 cap32; global8192 and all existing quotas remain unchanged. All four observed cells exercise specific changed addresses, values and readbacks; the generated matrix uses at most23/32 shared admissions. Existing native UI/rectangle and owning endpoints remain qualified. Post-host-fix matrix passes again.
The actual original receiver is0x09f22418, not the prior session's0x09f0115c. Packet87's member receiver matches the manager loaded through the native global in the owning slot0, endpoint1, guest thread1. Its instantiated observer SHA-256 is a6947c527a812d80a33d22894cc83abab2f55b8fd4ef45572047be54bb6d7fde. Runtime store packets prove the pending-delta handlers6ea32a/6ea339 set manager+c0/+c4 to(1,-145), then native block6ea77c writes+7c=255 and+78=401 with actual readbacks. Disassembly resolves those stores to6ea78a/6ea793 in6ea740. Block6e9775 clears the pending values. Threaded interpreter pointers in packets are not exact native instruction addresses.
The matching input singleton is0x074ffd1c. Its sampled+12c/+130 point is(401,255), with previous+134/+138=(400,300), while the trusted DOM(+1,-145) and host path record virtual(401,155). Thus the observed writer explains how a first member hit sees the cached(0,0) and a later hit sees a sampled point; the discrepancy already exists in the native input sample before these coordinate stores. The precise reason that native sample differs from the host point remains unresolved. This run does not prove the prior session's exact(400,251) production chain, and establishes no CPU discrepancy. Raw+a8 representation/name remains unknown and control+a4 is zero in the retained original evidence.
GetCursorPos retrieves a screen point and SetCursorPos moves it. Pointer Lock2.0 supplies motion deltas while locked. Our browser forwards those as guest counts; cursor warp and physical DirectInput motion are separate. A recentering software-cursor guest can legitimately receive button lParam(400,300) while its own UI accumulator hits(428,155). An absolute-looking locked driver point does not establish a guest UI location.
Original session
Exactly one original session accepted the EULA, displayed the intro/title, and rendered the Main Menu. Each input transition used a personally reviewed screenshot and the actual registered full shell. One center click acquired CANVAS pointer lock; member invocation200 received(0,0), returned AL0 and left the menu unchanged. No unchanged absolute clicks were repeated under lock.
After the native relative movement above, three smaller trusted Y moves(-35,-35,-30) were sent with waits. The bounded writer packets prove the first sample(401,220) and second sample(401,185). Shared kind24 reaches32/32 at17:59:17.138Z; no subsequent writer absence is inferred. The global observer did not saturate:3445 owning packets, maximum serial3442. The final balanced native click reaches member invocation241 at(401,185), returns AL0 and leaves Main Menu rendered in single-player-response.png. That point is below the retained Single Player hit rectangle's Y132..178. The rendered sword alone had been insufficient to establish the guest hit coordinate. No later writer state, missing move cause or target activation is inferred.
The ordinary-input queue's fixed deadline ended the browser at18:01:18UTC; the runner records code1 and ordinary-input queue deadline. The final browser exit is clean. A terminal read-only capture raced target closure and its error is retained. The exact first failing boundary is native menu targeting, before any Single Player transition. No mission loaded, gameplay movement or second distinct gameplay response was verified. Further work should trace how the owning native input sampler produces(401,255) from the proven host(401,155), and why the reviewed final click still samples(401,185), using a newly authorized session and the existing observer limits. No additional speculative production repair was made.