WINDOWS GRAB GESTURES, AUDIO AND FULLSCREEN REBUILD

ENEMY GRABS AND MOUSE CAPTURE
Donor-verified GameObjectPlayable::onEvent at 0x19b0b0 routes native tap 1001
and gesture 1007 to onGrappleEvent at 0x18b414 only in player state 9.
Its tap phase 5 does not read event coordinates. Native swipe events preserve
the original displacement/angle test. R1 uses emulateGrappleGestureEvent at
0x18b104, the game's own controller-compatible swipe routine, only in phase
1 or 7. No progress, health, timer or animation fields are overwritten.
The host maps fresh Cross/A presses, right-stick flicks and R1 to those paths;
it suspends ordinary movement/look/combat while grabbed. No raw pointer
down/move/up, mouse events or Android keys are fabricated by the controller.
The focused ABI fixture covers tap release/repress, flick rearm, held-input
phase transitions, paused/stale input, missing targets and VFP/FPSCR state.
The capture fixture covers menu/grab release, showing/ungrabbing the cursor,
resuming requested capture, explicit/focus-loss release and SDL failure state.
All 30 local regressions pass. See the download README for Windows startup,
fullscreen, audio and native-controller regression results for this EXE.
Enemy-grab sequences still require campaign verification on the user's PC;
the ordinary startup/controller run does not certify every QTE or enemy type.
Shake/motion-only interactions are not changed by this tap/swipe update.

WINDOWS AUDIO PACING
The diagnostic Windows run measured a synthetic 48 kHz producer at only
32795 frames/sec and 141 empty writes. The genuine mixer reported 48 kHz,
two channels and 512-frame blocks, but its stream measured 27632 frames/sec
with 584 empty writes during startup. No sample-rate mismatch was observed.
The revised submission-thread sleep uses accumulated deadlines and a Windows
high-resolution waitable timer. Normal non-audio guest sleeps are unchanged.
The original mixer/submit threads receive Windows HIGH priority. SDL playback
is prebuffered and starts with the original bounded queue limit.
The local clock fixture includes 200 ms of work in a one-second schedule and
checks that processing time does not accumulate into playback delay. Queue,
cancellation, mono/stereo reopen, donor rejection and hook idempotence pass.
It also injects two 45 ms stalls to exercise prompt low-water refilling.
The first Windows repair passed the timing and startup tests with no empty
writes, but the full software-rendered run exposed slow queue recovery.
Low-water refilling now uses a 1 ms yielding deadline until the queue recovers
its low-water mark. The subsequent long software-rendered run exposed gaps
up to 153 ms, exceeding the original 85 ms queue. Underruns now add bounded
headroom, capped at 12288 frames (256 ms at 48 kHz) and reduced after 60 stable
seconds. Unit checks cover growth, cap, decay and queued-PCM preservation.
The Windows check requires at least 90 percent of nominal PCM delivery and
at most eight queue underruns, including the full native-controller run.
Windows audible hardware quality is not established by dummy-audio tests.
See downloads/README.md for the current Windows timing and queue results.

DISPLAY
Windows startup now queries the desktop mode and defaults to borderless
fullscreen at that resolution instead of forcing 1366x768. SDL per-monitor
DPI awareness keeps drawable size and physical mouse coordinates aligned.
The Windows workflow checks native-resolution fullscreen on first start,
an F11 fullscreen/windowed round trip, and genuine nonblank frame captures.
The long software-rendered controller regression remains windowed at 1366x768.

The supplied Windows log confirms the original game was recompiled:
16475 translated/native functions were loaded, all 208 initializers completed,
JNI_OnLoad returned 0x10004, and the NVIDIA OpenGL window and WASAPI audio
device were created. The crash occurred after that initialization.

Reported exception: 0xc0000005, PC 0x00af2010, image base 0x00410000.
Normalizing to the executable's preferred base 0x00400000 gives 0x00ae2010.
The retained debug build resolves that instruction to f_00370bb4,
EA::core::GetAccelerometer(), original ARM RVA 0x00370bb4.
The failing instruction reads the null sensor object after device creation.

CAUSE AND FIX
linux_accel.h returned zero unconditionally on Windows. JNI consequently
reported no accelerometers. The original game still expects an accelerometer
and dereferences the failed device-creation result.
The Windows host now initializes a stationary logical device and returns one
sensor. SDL forwards rate-limited neutral gravity through the original native
acceleration callback only while the game enables acceleration.
No guest function, crash handler or null-pointer check is bypassed.

LOCAL CHECKS PASSED
Windows backend fixture: sensor count, stationary gravity, enable/disable,
all four delay categories, and timestamps above 32-bit tick range.
Existing real-IIO fixture: device selection, scaling, mounting, delay and errors.
Full Windows native game relink, retaining all translated game function objects.
Independent Win32 sensor test executable compiled for the Windows runner.

WINDOWS RUNTIME CHECK
The download branch contains a Windows-2022 workflow that tests this exact
package with the genuine upstream APK and test-only software OpenGL.
It runs both a fresh start and a second start with the resulting installation,
checks completion of 1200 frames per run, and captures logs and framebuffers.
See downloads/README.md on the download branch for its current result.
Test-only Mesa libraries and original game data are not in this ZIP.

NATIVE CONTROLLER CHECKS
Asset-free fixture passes native movement/look event layouts, unarmed melee
and stomp, equipped-weapon aim/fire guard, fire release, pause toggling, stale
input and disconnect, scene identity changes, live menu activation and hidden
or replaced widget rejection. Each callback preserves all VFP words and FPSCR.
Loading-screen confirm checks native loading completion before starting the
world and removing its loading layer; held buttons do not repeatedly continue.
Generator tests check donor rejection, hook placement, uniqueness/idempotence.
Static pad-path checks reject pointer, key or old swipe-helper dispatch.
The Windows workflow adds a virtual SDL pad run against the genuine game,
captures menu/gameplay frames and rejects legacy touch/key generation.
See downloads/README.md on the download branch for the actual runtime result.

LIMITS
No completed campaign, long-session stability, physical gamepad or motion
hardware validation is claimed. The native controller implementation is
reconstructed from the latest Linux binary, not recovered source code.
