Added First Person Camera Option

OPENXR.md

@@ -29,6 +29,11 @@ hud_width_meters = 2.4
 hud_virtual_screen = true
 stop_at_display_copy = true
 skip_copy_clears = true
+first_person = false
+first_person_units_per_meter = 10.0
+first_person_head_up_meters = 1.0
+first_person_head_forward_meters = 0.0
+first_person_head_right_meters = 0.0
 ```
 
 Set `enabled = true`, close the game completely, and start it again. These settings are read only
@@ -48,6 +53,40 @@ it is live and can be flipped from the F10 settings bar.
 `stop_at_display_copy` ends eye replay at the final `GXCopyDisp`, matching the frame shown on the
 desktop. `skip_copy_clears` independently suppresses the EFB reset performed after a copy. Both
 default on and can be changed live from the F10 settings bar for diagnostics.
+`first_person` and the `first_person_*` values are the first-person camera described below. All
+four are live and are also exposed in the F10 settings bar.
+
+## The first-person camera
+
+By default the headset sits where Mario Kart's own chase camera sits, and `world_units_per_meter`
+of 500 presents the race as a small diorama on a table. Turning on `first_person` moves the camera
+to the Player 1 driver's head instead, and switches the world scale to
+`first_person_units_per_meter`, which defaults to the 10 units per metre Mario Kart Wii is
+authored at, so the race reads life-size.
+
+The game's own transforms are never modified. Each guest frame the runtime reads the race camera's
+view matrix and the player kart's physics pose and derives one affine transform from the recorded
+view space into the space to render from. That transform is published with the sealed frame, and
+the renderer composes it onto every perspective draw's model-view matrix, alongside the headset's
+own per-eye delta. The kart's *physics* pose is used deliberately, not the animated model: an
+animated frame would bob and lurch the camera.
+
+Only the camera's heading is taken from the game. Its pitch and roll are dropped, so the horizon
+stays level through a chase-camera tilt or a banked corner, and the headset owns pitch, roll, and
+free look outright. The head's place in the kart is `first_person_head_up_meters` and its two
+companions, measured in the kart's own frame; the F10 sliders exist because the comfortable value
+is a matter of taste and is best judged from inside the headset.
+
+The mode engages only in a single-screen race, the same content that already qualifies for
+immersive stereo. Menus, split-screen, and the virtual-screen fallback are unaffected, and so is
+the desktop mirror, which keeps showing the game's ordinary third-person view. If the kart or
+camera cannot be read the camera stays where the game put it rather than guessing.
+
+Two limitations are worth knowing. Mario Kart still culls the scene from its own chase camera, so
+a wide head turn in first person can reveal the edge of what the game decided to draw. And the
+driver's own head is still rendered; nudge `first_person_head_forward_meters` if it intrudes. As
+with the rest of the race instrumentation, the object offsets this reads are specific to the
+project's supported PAL `RMCP01` translation.
 
 ## Presentation policy
 

aurora-main/include/aurora/aurora.h

@@ -87,10 +87,12 @@ enum { AURORA_STEREO_EYE_COUNT = 2 };
  * applies those four values to each perspective GX draw while preserving the
  * draw's own depth mapping and renderer depth-range adjustment.
  *
- * viewFromCenter is a row-major affine 3x4 transform from the game's
+ * viewFromCenter is a row-major affine 3x4 transform from the center-eye view
- * recorded center-eye view space into this eye's view space. Identity keeps
+ * space into this eye's view space. Identity keeps the recorded view and is
- * the recorded view and is useful when the game has already applied the eye
+ * useful when the game has already applied the eye transform before issuing GX
- * transform before issuing GX commands.
+ * commands. That center-eye space is the game's recorded view space unless
+ * aurora_set_stereo_scene_anchor() relocated the camera for the sealed frame,
+ * in which case the anchor is composed in for world draws only.
  *
  * Both transforms are ignored in AURORA_STEREO_FRAME_VIRTUAL_SCREEN mode.
  */
@@ -209,6 +211,24 @@ void aurora_end_frame();
 // Seal the current frame with an opaque application safety tag. Aurora rejects
 // an immersive provider packet unless its contentTag matches this exact frame.
 void aurora_end_frame_tagged(uint64_t contentTag);
+/**
+ * Relocates the immersive camera for the frame about to be sealed.
+ *
+ * anchorFromScene is a row-major affine 3x4 transform from the game's recorded
+ * view space into the view space the headset should render from, in world
+ * units. Identity keeps the recorded camera, which is the default and the
+ * behaviour of every frame that does not call this.
+ *
+ * Perspective draws carry the recorded camera in their own position matrices,
+ * so they are replayed through viewFromCenter * anchorFromScene. The 2D virtual
+ * screen is defined in the relocated camera's space and keeps viewFromCenter.
+ *
+ * This is latched by the next aurora_end_frame*(), then cleared: the anchor
+ * belongs to the guest frame that produced the GX content, so it must be
+ * published per frame from the producer thread rather than by the stereo
+ * provider, which cannot know which frame will consume its packet.
+ */
+void aurora_set_stereo_scene_anchor(const float anchorFromScene[12]);
 typedef void (*AuroraFrameWorkerWaitCallback)();
 // Called from the producer thread at bounded intervals while Aurora waits for
 // the asynchronous frame worker. The callback must not enter Aurora.

aurora-main/lib/aurora.cpp

@@ -3,6 +3,7 @@
 #ifdef AURORA_ENABLE_GX
 #include "gfx/common.hpp"
 #include "gfx/efb_ram_copy.hpp"
+#include "gfx/stereo_replay.hpp"
 #include "gx/fifo.hpp"
 #include "gx/shader_info.hpp"
 #include "imgui.hpp"
@@ -87,6 +88,22 @@ std::atomic_bool g_stereoProviderActive{false};
 StereoSinkRegistration g_stereoSink;
 #endif
 
+// First-person camera relocation for one sealed frame: a row-major affine 3x4
+// from the game's recorded view space into the space to render from. `active`
+// false means the identity transform, i.e. render from the recorded camera.
+struct StereoSceneAnchor {
+  std::array<float, 12> anchorFromScene{
+      1.f, 0.f, 0.f, 0.f, //
+      0.f, 1.f, 0.f, 0.f, //
+      0.f, 0.f, 1.f, 0.f,
+  };
+  bool active = false;
+};
+// Producer thread only, between aurora_set_stereo_scene_anchor() and the seal
+// that consumes it. Cleared at every seal so a producer that stops publishing
+// falls back to the recorded camera instead of freezing on a stale anchor.
+StereoSceneAnchor g_pendingSceneAnchor;
+
 using PresentClock = std::chrono::steady_clock;
 
 struct PresentTimingSample {
@@ -237,7 +254,7 @@ enum class ImGuiFramePolicy {
 bool begin_frame_impl(bool pumpEvents, ImGuiFramePolicy imguiPolicy = ImGuiFramePolicy::Immediate,
                       bool* imguiNewFrameOwed = nullptr) noexcept;
 bool begin_frame_render_state_impl(ImGuiFramePolicy imguiPolicy, bool* imguiNewFrameOwed) noexcept;
-void end_frame_impl(bool pumpEvents, bool drainFifo, uint64_t contentTag) noexcept;
+void end_frame_impl(bool pumpEvents, bool drainFifo, uint64_t contentTag, const StereoSceneAnchor& sceneAnchor) noexcept;
 
 // The two publication points of a frame-worker cycle, cleared together under `mutex`. Sealed:
 // producer-shared renderer state is free again. Done: slots encoded, presented, ImGui restarted.
@@ -254,9 +271,10 @@ struct FrameWorkerState {
   bool started = false;
   bool stop = false;
   bool jobPending = false;
-  // Written with jobPending and copied by the worker under this mutex. It
+  // Written with jobPending and copied by the worker under this mutex. They
-  // belongs to that exact queued frame, not to the producer's next frame.
+  // belong to that exact queued frame, not to the producer's next frame.
   uint64_t contentTag = AURORA_STEREO_CONTENT_TAG_UNKNOWN;
+  StereoSceneAnchor sceneAnchor{};
   // Readiness is polled thousands of times per frame, so these flags double as a publication
   // barrier. `sealed` is released before `ready`, and both are cleared under `mutex`.
   std::atomic_bool sealed{true};
@@ -305,7 +323,8 @@ bool frame_worker_requested() noexcept {
 
 #ifdef AURORA_ENABLE_GX
 // Returns false when a stop request was observed mid-cycle.
-bool run_frame_worker_cycle(gfx::SealedFrame& sealedFrame, uint64_t contentTag) noexcept;
+bool run_frame_worker_cycle(gfx::SealedFrame& sealedFrame, uint64_t contentTag,
+                            const StereoSceneAnchor& sceneAnchor) noexcept;
 #endif
 
 void frame_worker_main() noexcept {
@@ -322,6 +341,7 @@ void frame_worker_main() noexcept {
 
   for (;;) {
     uint64_t contentTag = AURORA_STEREO_CONTENT_TAG_UNKNOWN;
+    StereoSceneAnchor sceneAnchor{};
     {
       std::unique_lock lock(g_frameWorker.mutex);
       g_frameWorker.cv.wait(lock, [] { return g_frameWorker.stop || g_frameWorker.jobPending; });
@@ -330,17 +350,20 @@ void frame_worker_main() noexcept {
       }
       contentTag = g_frameWorker.contentTag;
       g_frameWorker.contentTag = AURORA_STEREO_CONTENT_TAG_UNKNOWN;
+      sceneAnchor = g_frameWorker.sceneAnchor;
+      g_frameWorker.sceneAnchor = {};
       g_frameWorker.jobPending = false;
     }
 
     // The CPU already decoded the sealed frame at its GX boundary; the worker only owns
     // encode/submit/present, so it never touches the producer's next FIFO buffer.
 #ifdef AURORA_ENABLE_GX
-    if (!run_frame_worker_cycle(sealedFrame, contentTag)) {
+    if (!run_frame_worker_cycle(sealedFrame, contentTag, sceneAnchor)) {
       break;
     }
 #else
     (void)contentTag;
+    (void)sceneAnchor;
 #endif
   }
 
@@ -365,6 +388,7 @@ void ensure_frame_worker_started() noexcept {
   g_frameWorker.stop = false;
   g_frameWorker.jobPending = false;
   g_frameWorker.contentTag = AURORA_STEREO_CONTENT_TAG_UNKNOWN;
+  g_frameWorker.sceneAnchor = {};
   g_frameWorker.sealed.store(true, std::memory_order_release);
   g_frameWorker.ready.store(true, std::memory_order_release);
   g_frameWorker.prepareAllowed = false;
@@ -433,6 +457,7 @@ void stop_frame_worker() noexcept {
   g_frameWorker.framePrepared = false;
   g_frameWorker.jobPending = false;
   g_frameWorker.contentTag = AURORA_STEREO_CONTENT_TAG_UNKNOWN;
+  g_frameWorker.sceneAnchor = {};
 }
 
 uint32_t align_to(uint32_t value, uint32_t alignment) noexcept { return (value + alignment - 1) & ~(alignment - 1); }
@@ -608,7 +633,9 @@ std::optional<AuroraStereoFrame> request_stereo_frame(uint32_t logicalFrame, uin
   return frame;
 }
 
-gfx::StereoReplayFrame make_stereo_replay_frame(const AuroraStereoFrame& input) {
+gfx::StereoReplayFrame make_stereo_replay_frame(const AuroraStereoFrame& input, const StereoSceneAnchor& sceneAnchor) {
+  Mat3x4<float> anchorFromScene;
+  std::memcpy(&anchorFromScene, sceneAnchor.anchorFromScene.data(), sizeof(anchorFromScene));
   gfx::StereoReplayFrame replay{};
   for (uint32_t eye = 0; eye < AURORA_STEREO_EYE_COUNT; ++eye) {
     ensure_stereo_eye_target(eye, input.eyes[eye].width, input.eyes[eye].height);
@@ -627,6 +654,12 @@ gfx::StereoReplayFrame make_stereo_replay_frame(const AuroraStereoFrame& input)
     };
     std::memcpy(&view.projection, input.eyes[eye].projection, sizeof(view.projection));
     std::memcpy(&view.viewFromCenter, input.eyes[eye].viewFromCenter, sizeof(view.viewFromCenter));
+    // World draws already carry the recorded camera, so they need the anchor
+    // folded in; the virtual screen is authored in the anchored camera's space
+    // and keeps viewFromCenter.
+    view.viewFromScene = sceneAnchor.active
+                             ? gfx::stereo_replay::compose_affine(view.viewFromCenter, anchorFromScene)
+                             : view.viewFromCenter;
   }
   return replay;
 }
@@ -1556,7 +1589,8 @@ struct SealedFrameContext {
 
 // Phase 1: everything that touches producer-shared renderer state. Needs g_rendererGpuMutex and
 // a FIFO already drained into the recorded pass list.
-void seal_frame_locked(gfx::SealedFrame& sealedFrame, SealedFrameContext& ctx, uint64_t contentTag) {
+void seal_frame_locked(gfx::SealedFrame& sealedFrame, SealedFrameContext& ctx, uint64_t contentTag,
+                       const StereoSceneAnchor& sceneAnchor) {
   ZoneScopedN("Seal frame");
   const auto encoderDescriptor = wgpu::CommandEncoderDescriptor{
       .label = "Redraw encoder",
@@ -1571,7 +1605,7 @@ void seal_frame_locked(gfx::SealedFrame& sealedFrame, SealedFrameContext& ctx, u
   if (const auto stereoInput = request_stereo_frame(ctx.logicalFrame, contentTag)) {
     ctx.stereoFrameToken = stereoInput->frameToken;
     ctx.stereoFrameMode = stereoInput->mode;
-    ctx.stereoReplay = make_stereo_replay_frame(*stereoInput);
+    ctx.stereoReplay = make_stereo_replay_frame(*stereoInput, sceneAnchor);
   }
   if (ctx.stereoReplay && ctx.stereoFrameMode == AURORA_STEREO_FRAME_IMMERSIVE_REPLAY) {
     // Keep the accepted frame alive even if the additional eye-uniform copies
@@ -1852,7 +1886,8 @@ void record_frame_telemetry() {
 
 // One complete frame-worker cycle. The scene encode only leaves the renderer mutex when
 // interpolation actually inserts slots; otherwise both phases publish together.
-bool run_frame_worker_cycle(gfx::SealedFrame& sealedFrame, uint64_t contentTag) noexcept {
+bool run_frame_worker_cycle(gfx::SealedFrame& sealedFrame, uint64_t contentTag,
+                            const StereoSceneAnchor& sceneAnchor) noexcept {
   ZoneScopedN("Frame worker cycle");
   webgpu::fail_if_device_lost();
   SealedFrameContext ctx;
@@ -1860,7 +1895,7 @@ bool run_frame_worker_cycle(gfx::SealedFrame& sealedFrame, uint64_t contentTag)
   bool overlapEncode = false;
   {
     std::lock_guard gpuLock(g_rendererGpuMutex);
-    seal_frame_locked(sealedFrame, ctx, contentTag);
+    seal_frame_locked(sealedFrame, ctx, contentTag, sceneAnchor);
     overlapEncode = ctx.interpolationActive;
     if (!overlapEncode) {
       presentationJobs = encode_sealed_frame(sealedFrame, ctx);
@@ -1919,7 +1954,8 @@ bool run_frame_worker_cycle(gfx::SealedFrame& sealedFrame, uint64_t contentTag)
 
 // Synchronous frame submission: seal, encode and present inline on the calling thread. Used when
 // the frame worker is disabled (RenderDoc captures) and on the boot path.
-void end_frame_impl(bool pumpEvents, bool drainFifo, uint64_t contentTag) noexcept {
+void end_frame_impl(bool pumpEvents, bool drainFifo, uint64_t contentTag,
+                    const StereoSceneAnchor& sceneAnchor) noexcept {
   ZoneScoped;
 #ifdef AURORA_ENABLE_GX
   webgpu::fail_if_device_lost();
@@ -1934,7 +1970,7 @@ void end_frame_impl(bool pumpEvents, bool drainFifo, uint64_t contentTag) noexce
     if (drainFifo) {
       gx::fifo::drain();
     }
-    seal_frame_locked(sealedFrame, ctx, contentTag);
+    seal_frame_locked(sealedFrame, ctx, contentTag, sceneAnchor);
     presentationJobs = encode_sealed_frame(sealedFrame, ctx);
   }
   publish_presentations(std::move(presentationJobs), ctx.interpolationActive);
@@ -1943,6 +1979,7 @@ void end_frame_impl(bool pumpEvents, bool drainFifo, uint64_t contentTag) noexce
   (void)pumpEvents;
   (void)drainFifo;
   (void)contentTag;
+  (void)sceneAnchor;
 #endif
 }
 
@@ -2013,8 +2050,12 @@ void end_frame(uint64_t contentTag) noexcept {
 #ifdef AURORA_ENABLE_GX
   webgpu::fail_if_device_lost();
 #endif
+  // Claim the anchor published for this frame. Clearing it here is what makes a
+  // producer that stops publishing fall back to the recorded camera.
+  const StereoSceneAnchor sceneAnchor = g_pendingSceneAnchor;
+  g_pendingSceneAnchor = {};
   if (!frame_worker_requested()) {
-    end_frame_impl(true, true, contentTag);
+    end_frame_impl(true, true, contentTag, sceneAnchor);
     return;
   }
 
@@ -2035,6 +2076,7 @@ void end_frame(uint64_t contentTag) noexcept {
     g_frameWorker.sealed.store(false, std::memory_order_release);
     g_frameWorker.ready.store(false, std::memory_order_release);
     g_frameWorker.contentTag = contentTag;
+    g_frameWorker.sceneAnchor = sceneAnchor;
     g_frameWorker.jobPending = true;
     g_frameWorker.prepareAllowed = false;
   }
@@ -2085,6 +2127,32 @@ void set_stereo_frame_provider(AuroraStereoFrameProvider provider, void* userdat
   g_stereoProviderActive.store(provider != nullptr, std::memory_order_release);
 }
 
+void set_stereo_scene_anchor(const float anchorFromScene[12]) noexcept {
+  if (anchorFromScene == nullptr) {
+    g_pendingSceneAnchor = {};
+    return;
+  }
+  // aurora_core is built with -ffast-math, so std::isfinite may be folded to
+  // true. Inspect the IEEE-754 exponent, matching request_stereo_frame().
+  for (size_t i = 0; i < 12; ++i) {
+    uint32_t bits = 0;
+    std::memcpy(&bits, &anchorFromScene[i], sizeof(bits));
+    if ((bits & 0x7f800000u) == 0x7f800000u) {
+      static bool rejectionLogged = false;
+      if (!rejectionLogged) {
+        rejectionLogged = true;
+        Log.warn("Rejected a non-finite stereo scene anchor; keeping the recorded camera");
+      }
+      g_pendingSceneAnchor = {};
+      return;
+    }
+  }
+  StereoSceneAnchor anchor{};
+  std::memcpy(anchor.anchorFromScene.data(), anchorFromScene, sizeof(anchor.anchorFromScene));
+  anchor.active = true;
+  g_pendingSceneAnchor = anchor;
+}
+
 #ifdef AURORA_ENABLE_GX
 namespace stereo {
 void set_sink(SinkCallback callback, SubmitCallback submitted, void* userdata) noexcept {
@@ -2108,6 +2176,9 @@ const AuroraEvent* aurora_update() { return aurora::update(); }
 bool aurora_begin_frame() { return aurora::begin_frame(); }
 void aurora_end_frame() { aurora::end_frame(AURORA_STEREO_CONTENT_TAG_UNKNOWN); }
 void aurora_end_frame_tagged(uint64_t contentTag) { aurora::end_frame(contentTag); }
+void aurora_set_stereo_scene_anchor(const float anchorFromScene[12]) {
+  aurora::set_stereo_scene_anchor(anchorFromScene);
+}
 void aurora_set_frame_worker_wait_callback(AuroraFrameWorkerWaitCallback callback) {
   aurora::g_frameWorkerWaitCallback.store(callback, std::memory_order_release);
 }

aurora-main/lib/gfx/common.cpp

@@ -1366,19 +1366,22 @@ static bool prepare_stereo_replay_uniforms(const StereoReplayFrame& stereoFrame)
             const size_t offset = layout.positionOffset + matrix * sizeof(Mat3x4<float>);
             Mat3x4<float> source;
             std::memcpy(&source, uniform.data() + offset, sizeof(source));
-            const auto transformed = stereo_replay::compose_affine(eye.viewFromCenter, source);
+            const auto transformed = stereo_replay::compose_affine(eye.viewFromScene, source);
             std::memcpy(uniform.data() + offset, &transformed, sizeof(transformed));
           }
           for (uint32_t matrix = 0; matrix < layout.normalMatrixCount; ++matrix) {
             const size_t offset = layout.normalOffset + matrix * sizeof(Mat3x4<float>);
             Mat3x4<float> source;
             std::memcpy(&source, uniform.data() + offset, sizeof(source));
-            const auto transformed = stereo_replay::compose_normal(eye.viewFromCenter, source);
+            const auto transformed = stereo_replay::compose_normal(eye.viewFromScene, source);
             std::memcpy(uniform.data() + offset, &transformed, sizeof(transformed));
           }
         } else {
           // 2D content reaches the eye entirely through its projection: the
           // draw's own position matrices lay the element out in screen space.
+          // The screen rectangle is built in the VR-neutral view space, so this
+          // path uses viewFromCenter, not viewFromScene: folding the anchor in
+          // would leave the screen behind at the camera the anchor replaced.
           // First lift viewport-local NDC into displayed-frame NDC; replay will
           // use a full-eye viewport so sub-pane elements are not transformed by
           // the recorded viewport a second time.

aurora-main/lib/gfx/common.hpp

@@ -295,7 +295,15 @@ struct ReplayTarget {
 struct StereoReplayEye {
   ReplayTarget target;
   Mat4x4<float> projection;
+  // The headset's eye delta, from the VR-neutral view space into this eye's.
+  // The virtual screen is defined in that neutral space, so 2D reprojection
+  // uses this transform directly.
   Mat3x4<float> viewFromCenter;
+  // The same delta with the first-person scene anchor folded in, i.e. from the
+  // game's *recorded* view space into this eye's. World draws carry the game's
+  // camera in their position matrices and therefore need this one. It equals
+  // viewFromCenter whenever the anchor is identity.
+  Mat3x4<float> viewFromScene;
 };
 
 struct StereoReplayFrame {

aurora-main/tests/stereo_replay_test.cpp

@@ -155,5 +155,101 @@ TEST(StereoReplayTest, HudScreenParksRasterDepthAtMidrangeUnderHeadMotion) {
   }
 }
 
+Mat3x4<float> identity3x4() {
+  Mat3x4<float> m{};
+  m.m0 = {1.0f, 0.0f, 0.0f, 0.0f};
+  m.m1 = {0.0f, 1.0f, 0.0f, 0.0f};
+  m.m2 = {0.0f, 0.0f, 1.0f, 0.0f};
+  return m;
+}
+
+Mat3x4<float> head_tracking_delta() {
+  const float angle = 0.21f;
+  const float c = std::cos(angle);
+  const float s = std::sin(angle);
+  Mat3x4<float> m{};
+  m.m0 = {c, 0.0f, s, 11.0f};
+  m.m1 = {0.0f, 1.0f, 0.0f, -3.0f};
+  m.m2 = {-s, 0.0f, c, 6.0f};
+  return m;
+}
+
+TEST(StereoReplayTest, IdentitySceneAnchorLeavesTheEyeDeltaUnchanged) {
+  const auto viewFromCenter = head_tracking_delta();
+
+  const auto viewFromScene = compose_affine(viewFromCenter, identity3x4());
+
+  EXPECT_EQ(viewFromScene, viewFromCenter);
+}
+
+TEST(StereoReplayTest, TranslatingSceneAnchorMovesTheWorldByTheAnchorOffset) {
+  // A first-person anchor with no levelling is translate(-a): the camera moves
+  // to a, so every world point must arrive a units closer to the eye origin.
+  const std::array<float, 3> a{40.0f, -12.0f, -260.0f};
+  auto anchor = identity3x4();
+  anchor.m0[3] = -a[0];
+  anchor.m1[3] = -a[1];
+  anchor.m2[3] = -a[2];
+  const auto viewFromCenter = head_tracking_delta();
+  const auto viewFromScene = compose_affine(viewFromCenter, anchor);
+
+  // An object matrix placing a vertex somewhere in the recorded view space.
+  Mat3x4<float> objectToCenter{};
+  objectToCenter.m0 = {1.0f, 0.0f, 0.0f, 130.0f};
+  objectToCenter.m1 = {0.0f, 1.0f, 0.0f, 55.0f};
+  objectToCenter.m2 = {0.0f, 0.0f, 1.0f, -900.0f};
+
+  const auto anchored = compose_affine(viewFromScene, objectToCenter);
+  const auto recorded = compose_affine(viewFromCenter, objectToCenter);
+
+  // Rotation is untouched, and the eye-space displacement is exactly the eye
+  // delta's rotation applied to -a.
+  for (size_t row = 0; row < 3; ++row) {
+    const auto& anchoredRow = *(&anchored.m0 + row);
+    const auto& recordedRow = *(&recorded.m0 + row);
+    const auto& viewRow = *(&viewFromCenter.m0 + row);
+    for (size_t column = 0; column < 3; ++column) {
+      EXPECT_FLOAT_EQ(anchoredRow[column], recordedRow[column]);
+    }
+    const float expected =
+        recordedRow[3] - (viewRow[0] * a[0] + viewRow[1] * a[1] + viewRow[2] * a[2]);
+    EXPECT_NEAR(anchoredRow[3], expected, 1e-3f);
+  }
+}
+
+TEST(StereoReplayTest, VirtualScreenStaysAheadOfTheAnchoredCamera) {
+  // The screen rectangle is authored in the anchored camera's space and so
+  // composes with viewFromCenter, while world geometry composes with
+  // viewFromScene. The two agree exactly when a world object placed `distance`
+  // ahead of the anchored camera lands on the screen's centre.
+  const std::array<float, 3> a{40.0f, -12.0f, -260.0f};
+  const float distance = 20.0f;
+  auto anchor = identity3x4();
+  anchor.m0[3] = -a[0];
+  anchor.m1[3] = -a[1];
+  anchor.m2[3] = -a[2];
+  const auto viewFromCenter = head_tracking_delta();
+  const auto viewFromScene = compose_affine(viewFromCenter, anchor);
+
+  // The screen's centre: (0, 0, -distance) in the anchored camera's space,
+  // carried into eye space by viewFromCenter alone.
+  const Vec4<float> screenCentre{0.0f, 0.0f, -distance, 1.0f};
+  const float centreX = dot4(viewFromCenter.m0, screenCentre);
+  const float centreY = dot4(viewFromCenter.m1, screenCentre);
+  const float centreZ = dot4(viewFromCenter.m2, screenCentre);
+
+  // A world object at the same place, expressed the way a GX draw carries it:
+  // in the *recorded* view space, hence offset by the anchor position.
+  Mat3x4<float> objectToCenter = identity3x4();
+  objectToCenter.m0[3] = a[0];
+  objectToCenter.m1[3] = a[1];
+  objectToCenter.m2[3] = a[2] - distance;
+  const auto placed = compose_affine(viewFromScene, objectToCenter);
+
+  EXPECT_NEAR(placed.m0[3], centreX, 1e-3f);
+  EXPECT_NEAR(placed.m1[3], centreY, 1e-3f);
+  EXPECT_NEAR(placed.m2[3], centreZ, 1e-3f);
+}
+
 } // namespace
 } // namespace aurora::gfx::stereo_replay

runtime/CMakeLists.txt

@@ -335,6 +335,13 @@ target_link_libraries(mkw_platform_paths_tests PRIVATE mkw_platform)
 target_compile_features(mkw_platform_paths_tests PRIVATE cxx_std_17)
 add_test(NAME mkw_platform_paths_tests COMMAND mkw_platform_paths_tests)
 
+# The first-person VR camera's transform is deliberately header-only and free of
+# guest access so it can be checked here, without a headset or a running game.
+add_executable(mkw_vr_first_person_tests "${CMAKE_CURRENT_LIST_DIR}/tests/vr_first_person_tests.cpp")
+target_include_directories(mkw_vr_first_person_tests PRIVATE "${CMAKE_CURRENT_LIST_DIR}/include")
+target_compile_features(mkw_vr_first_person_tests PRIVATE cxx_std_17)
+add_test(NAME mkw_vr_first_person_tests COMMAND mkw_vr_first_person_tests)
+
 # HostContext deliberately keeps the platform-specific context primitive out
 # of fiber_manager.cpp. Exercise the Linux libco handoff directly so future
 # refactors cannot silently remove its headers, implementation, or link edge.

runtime/include/runtime_config.h

@@ -55,6 +55,11 @@ struct RuntimeUserConfig {
     std::optional<bool> vrHudVirtualScreen;
     std::optional<bool> vrStopAtDisplayCopy;
     std::optional<bool> vrSkipCopyClears;
+    std::optional<bool> vrFirstPerson;
+    std::optional<float> vrFirstPersonUnitsPerMeter;
+    std::optional<float> vrFirstPersonHeadUpMeters;
+    std::optional<float> vrFirstPersonHeadForwardMeters;
+    std::optional<float> vrFirstPersonHeadRightMeters;
     std::optional<float> audioVolume;
     std::optional<float> audioMusicVolume;
     std::optional<float> audioSoundEffectsVolume;
@@ -339,7 +344,18 @@ inline void EnsureConfigFile() {
               "# reset a GX copy performs afterwards. Both keep that reset\n"
               "# from erasing the eye, and both are safe to turn off.\n"
               "stop_at_display_copy = true\n"
-              "skip_copy_clears = true\n\n"
+              "skip_copy_clears = true\n"
+              "# Put the camera at the Player 1 driver's head instead of behind\n"
+              "# the kart, with the horizon kept level. Changeable live from the\n"
+              "# F10 menu, and only during a single-screen race. The world scale\n"
+              "# below replaces world_units_per_meter while it is engaged: 10 is\n"
+              "# life-size, where the 500 above makes the race a small diorama.\n"
+              "first_person = false\n"
+              "first_person_units_per_meter = 10.0\n"
+              "# Where the head sits in the kart's own frame, in metres.\n"
+              "first_person_head_up_meters = 1.0\n"
+              "first_person_head_forward_meters = 12.0\n"
+              "first_person_head_right_meters = 0.0\n\n"
               "[audio]\n"
               "volume = 1.0\n"
               "music_volume = 1.0\n"
@@ -500,6 +516,23 @@ inline RuntimeUserConfig ParseConfigDocument(const toml::value& document) {
     config.vrHudVirtualScreen = FindConfigValue<bool>(document, "vr", "hud_virtual_screen");
     config.vrStopAtDisplayCopy = FindConfigValue<bool>(document, "vr", "stop_at_display_copy");
     config.vrSkipCopyClears = FindConfigValue<bool>(document, "vr", "skip_copy_clears");
+    config.vrFirstPerson = FindConfigValue<bool>(document, "vr", "first_person");
+    if (auto value = FindConfigFloat(document, "vr", "first_person_units_per_meter");
+        value && *value >= 1.0f && *value <= 10000.0f) {
+        config.vrFirstPersonUnitsPerMeter = *value;
+    }
+    if (auto value = FindConfigFloat(document, "vr", "first_person_head_up_meters");
+        value && *value >= -3.0f && *value <= 3.0f) {
+        config.vrFirstPersonHeadUpMeters = *value;
+    }
+    if (auto value = FindConfigFloat(document, "vr", "first_person_head_forward_meters");
+        value && *value >= -3.0f && *value <= 3.0f) {
+        config.vrFirstPersonHeadForwardMeters = *value;
+    }
+    if (auto value = FindConfigFloat(document, "vr", "first_person_head_right_meters");
+        value && *value >= -3.0f && *value <= 3.0f) {
+        config.vrFirstPersonHeadRightMeters = *value;
+    }
 
     auto readVolume = [&](std::string_view key) -> std::optional<float> {
         auto value = FindConfigFloat(document, "audio", key);
@@ -742,6 +775,43 @@ inline bool SetVrSkipCopyClears(bool value) {
     return WriteSetting("vr", "skip_copy_clears", value ? "true" : "false");
 }
 
+inline bool SetVrFirstPerson(bool value) {
+    Mutable().vrFirstPerson = value;
+    return WriteSetting("vr", "first_person", value ? "true" : "false");
+}
+
+inline bool SetVrFirstPersonUnitsPerMeter(float value) {
+    value = std::clamp(value, 1.0f, 10000.0f);
+    Mutable().vrFirstPersonUnitsPerMeter = value;
+    std::ostringstream formatted;
+    formatted << value;
+    return WriteSetting("vr", "first_person_units_per_meter", formatted.str());
+}
+
+inline bool SetVrFirstPersonHeadUpMeters(float value) {
+    value = std::clamp(value, -3.0f, 3.0f);
+    Mutable().vrFirstPersonHeadUpMeters = value;
+    std::ostringstream formatted;
+    formatted << value;
+    return WriteSetting("vr", "first_person_head_up_meters", formatted.str());
+}
+
+inline bool SetVrFirstPersonHeadForwardMeters(float value) {
+    value = std::clamp(value, -3.0f, 3.0f);
+    Mutable().vrFirstPersonHeadForwardMeters = value;
+    std::ostringstream formatted;
+    formatted << value;
+    return WriteSetting("vr", "first_person_head_forward_meters", formatted.str());
+}
+
+inline bool SetVrFirstPersonHeadRightMeters(float value) {
+    value = std::clamp(value, -3.0f, 3.0f);
+    Mutable().vrFirstPersonHeadRightMeters = value;
+    std::ostringstream formatted;
+    formatted << value;
+    return WriteSetting("vr", "first_person_head_right_meters", formatted.str());
+}
+
 inline bool SetControllerButton(size_t index, std::string value) {
     if (index >= kControllerButtonKeys.size()) {
         return false;
@@ -990,6 +1060,26 @@ inline bool VrSkipCopyClears(bool fallback = true) {
     return Get().vrSkipCopyClears.value_or(fallback);
 }
 
+inline bool VrFirstPerson(bool fallback = false) {
+    return Get().vrFirstPerson.value_or(fallback);
+}
+
+inline float VrFirstPersonUnitsPerMeter(float fallback = 10.0f) {
+    return std::clamp(Get().vrFirstPersonUnitsPerMeter.value_or(fallback), 1.0f, 10000.0f);
+}
+
+inline float VrFirstPersonHeadUpMeters(float fallback = 1.0f) {
+    return std::clamp(Get().vrFirstPersonHeadUpMeters.value_or(fallback), -3.0f, 3.0f);
+}
+
+inline float VrFirstPersonHeadForwardMeters(float fallback = 0.0f) {
+    return std::clamp(Get().vrFirstPersonHeadForwardMeters.value_or(fallback), -20.0f, 20.0f);
+}
+
+inline float VrFirstPersonHeadRightMeters(float fallback = 0.0f) {
+    return std::clamp(Get().vrFirstPersonHeadRightMeters.value_or(fallback), -3.0f, 3.0f);
+}
+
 inline std::string GraphicsApi(std::string fallback = "auto") {
     return Get().graphicsApi.value_or(std::move(fallback));
 }

runtime/include/settings_overlay.h

@@ -12,4 +12,8 @@ void Draw() noexcept;
 bool StartupScreenVisible() noexcept;
 void NotifyStrapInputAccepted() noexcept;
 void AdvancePresentedFrame() noexcept;
+// Re-sends the VR virtual screen's placement to Aurora. Its metres are
+// converted with the world scale currently in effect, so switching the
+// first-person camera on or off has to repeat it.
+void RefreshVrHudVirtualScreen() noexcept;
 } // namespace settings_overlay

runtime/include/vr/mkw_vr_first_person.h

@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+
+namespace mkw::vr {
+
+// A row-major affine 3x4, the same shape and convention as an NW4R/GX Mtx and
+// as Aurora's Mat3x4: a point is transformed as out = M * (p, 1).
+using Mtx34 = std::array<float, 12>;
+
+inline constexpr Mtx34 kIdentityMtx34{
+    1.0f, 0.0f, 0.0f, 0.0f, //
+    0.0f, 1.0f, 0.0f, 0.0f, //
+    0.0f, 0.0f, 1.0f, 0.0f,
+};
+
+// Where the driver's head sits in the kart's own frame, in metres. The kart
+// frame is the EGG convention: +x right, +y up, +z forward.
+struct FirstPersonHeadOffsets {
+    float right = 0.0f;
+    float up = 1.0f;
+    float forward = 0.0f;
+};
+
+// The camera relocation published to Aurora for one guest frame: a transform
+// from the game's recorded view space into the space the headset renders from.
+struct FirstPersonAnchor {
+    Mtx34 anchor_from_scene = kIdentityMtx34;
+    bool valid = false;
+    uint64_t guest_frame_index = 0;
+};
+
+// ---------------------------------------------------------------------------
+// Pure math. Header-only and free of guest access, so it is directly testable.
+// ---------------------------------------------------------------------------
+
+namespace detail {
+
+inline constexpr float kAnchorEpsilon = 1.0e-6f;
+
+inline bool IsFiniteFloat(const float* value) noexcept {
+    // The runtime is built with -ffast-math, which permits the compiler to fold
+    // std::isfinite to true. Inspect the object representation instead, the way
+    // the presentation policy validates its own floats.
+    uint32_t bits = 0;
+    std::memcpy(&bits, value, sizeof(bits));
+    return (bits & 0x7F800000u) != 0x7F800000u;
+}
+
+inline bool IsFiniteMtx34(const Mtx34& value) noexcept {
+    for (const float& element : value) {
+        if (!IsFiniteFloat(&element)) {
+            return false;
+        }
+    }
+    return true;
+}
+
+struct Vec3 {
+    float x = 0.0f;
+    float y = 0.0f;
+    float z = 0.0f;
+};
+
+inline float Dot(const Vec3& a, const Vec3& b) noexcept {
+    return a.x * b.x + a.y * b.y + a.z * b.z;
+}
+
+inline Vec3 Cross(const Vec3& a, const Vec3& b) noexcept {
+    return {a.y * b.z - a.z * b.y, a.z * b.x - a.x * b.z, a.x * b.y - a.y * b.x};
+}
+
+inline bool Normalize(Vec3& value) noexcept {
+    const float length_squared = Dot(value, value);
+    if (!IsFiniteFloat(&length_squared) || !(length_squared > kAnchorEpsilon)) {
+        return false;
+    }
+    const float inverse_length = 1.0f / std::sqrt(length_squared);
+    value.x *= inverse_length;
+    value.y *= inverse_length;
+    value.z *= inverse_length;
+    return true;
+}
+
+// out = matrix * (x, y, z, 1)
+inline Vec3 TransformPoint(const Mtx34& matrix, float x, float y, float z) noexcept {
+    return {
+        matrix[0] * x + matrix[1] * y + matrix[2] * z + matrix[3],
+        matrix[4] * x + matrix[5] * y + matrix[6] * z + matrix[7],
+        matrix[8] * x + matrix[9] * y + matrix[10] * z + matrix[11],
+    };
+}
+
+} // namespace detail
+
+// Builds the anchor from the game's view matrix (world -> recorded view space),
+// the kart's pose (kart-local -> world), and head offsets already converted to
+// world units.
+//
+// The translation moves the camera onto the head. With level_horizon the
+// rotation keeps the recorded camera's heading but drops its pitch and roll, so
+// the headset owns pitch and roll outright; without it the recorded camera's
+// orientation is kept whole and only the eye moves. Returns false and leaves
+// `out` untouched when the inputs cannot produce an orthonormal frame.
+inline bool ComputeFirstPersonAnchor(const Mtx34& view_from_world, const Mtx34& kart_from_local,
+                                     float head_right_units, float head_up_units,
+                                     float head_forward_units, bool level_horizon,
+                                     Mtx34& out) noexcept {
+    using namespace detail;
+    if (!IsFiniteMtx34(view_from_world) || !IsFiniteMtx34(kart_from_local)) {
+        return false;
+    }
+    const Vec3 head_world =
+        TransformPoint(kart_from_local, head_right_units, head_up_units, head_forward_units);
+    const Vec3 a = TransformPoint(view_from_world, head_world.x, head_world.y, head_world.z);
+    if (!IsFiniteFloat(&a.x) || !IsFiniteFloat(&a.y) || !IsFiniteFloat(&a.z)) {
+        return false;
+    }
+
+    // Rows of the anchor's rotation. Identity keeps the recorded camera's own
+    // orientation and moves the eye only.
+    Vec3 rows[3]{{1.0f, 0.0f, 0.0f}, {0.0f, 1.0f, 0.0f}, {0.0f, 0.0f, 1.0f}};
+    if (level_horizon) {
+        // World +Y in view coordinates: the column of the view rotation that
+        // the world up axis selects.
+        Vec3 up{view_from_world[1], view_from_world[5], view_from_world[9]};
+        if (!Normalize(up)) {
+            return false;
+        }
+        // Level the recorded camera's forward (-Z in its own space) onto the
+        // horizon plane. Looking near-straight up or down leaves nothing to
+        // project, so recover the heading from the camera's up axis instead.
+        const Vec3 camera_forward{0.0f, 0.0f, -1.0f};
+        float along = Dot(camera_forward, up);
+        Vec3 forward{camera_forward.x - up.x * along, camera_forward.y - up.y * along,
+                     camera_forward.z - up.z * along};
+        if (!Normalize(forward)) {
+            const Vec3 camera_up{0.0f, 1.0f, 0.0f};
+            along = Dot(camera_up, up);
+            forward = {camera_up.x - up.x * along, camera_up.y - up.y * along,
+                       camera_up.z - up.z * along};
+            if (!Normalize(forward)) {
+                return false;
+            }
+        }
+        Vec3 right = Cross(forward, up);
+        if (!Normalize(right)) {
+            return false;
+        }
+        // Re-derive up from the orthonormalized pair so a slightly non-rigid
+        // view matrix cannot leave a skewed frame behind.
+        rows[0] = right;
+        rows[1] = Cross(right, forward);
+        rows[2] = {-forward.x, -forward.y, -forward.z};
+    }
+
+    Mtx34 anchor{};
+    for (uint32_t row = 0; row < 3; ++row) {
+        anchor[row * 4 + 0] = rows[row].x;
+        anchor[row * 4 + 1] = rows[row].y;
+        anchor[row * 4 + 2] = rows[row].z;
+        anchor[row * 4 + 3] = -Dot(rows[row], a);
+    }
+    if (!IsFiniteMtx34(anchor)) {
+        return false;
+    }
+    out = anchor;
+    return true;
+}
+
+// ---------------------------------------------------------------------------
+// Per-frame observation. Called from the translated-code observers on the guest
+// thread; the anchor is consumed by the producer at its Aurora frame seal.
+// ---------------------------------------------------------------------------
+
+// Enables anchor computation and sets the head offsets and world scale used to
+// convert them. Called whenever the configuration or the F10 toggle changes.
+void MkwVRFirstPersonConfigure(bool enabled, const FirstPersonHeadOffsets& offsets,
+                               float units_per_meter) noexcept;
+
+// Reads the current [vr] first-person settings and applies them here and to the
+// presentation policy's world scale. The single place those settings are
+// interpreted, shared by startup and the F10 settings bar.
+void MkwVRFirstPersonApplyConfiguredSettings() noexcept;
+
+// Reads the race camera and the player's kart and republishes the anchor. Call
+// once per guest frame, after the kart and camera updates and before the draws.
+// race_camera_address is the frame's own RaceCamera, or zero if none was seen.
+void MkwVRFirstPersonUpdate(uint64_t guest_frame_index, uint32_t race_camera_address) noexcept;
+
+// Drops every captured pointer and the held anchor. Call on race entry/exit.
+void MkwVRFirstPersonReset() noexcept;
+
+// Producer-side read. Thread-safe. A valid anchor is also what marks the mode
+// as engaged, and so what selects the first-person world scale: it is invalid
+// whenever the mode is off, the race has not produced a usable anchor, or the
+// anchor has been missing long enough to give up holding the last one.
+FirstPersonAnchor MkwVRFirstPersonGetAnchor() noexcept;
+
+} // namespace mkw::vr

runtime/include/vr/mkw_vr_policy.h

@@ -78,6 +78,11 @@ struct MkwVRPolicyConfig {
     // immersive race HUD so 2D content keeps its place across the transition.
     float hud_width_meters = 2.4f;
     float hud_scale = 1.0f;
+    // World scale used while the first-person camera is engaged. Mario Kart
+    // Wii is authored at roughly this many units per metre, so it is what
+    // makes the race read life-size; the third-person default deliberately
+    // does not, presenting the race as a small diorama instead.
+    float first_person_units_per_meter = 10.0f;
 };
 
 struct MkwVRSceneObservation {
@@ -108,6 +113,10 @@ struct MkwVRPolicySnapshot {
     MkwVRCameraObservation camera{};
     uint32_t available_bindings = MkwVRBindingNone;
     bool session_active = false;
+    // A first-person camera is actually driving the view this frame. Set by the
+    // integration layer once the anchor it publishes to the renderer is valid,
+    // so the world scale can never disagree with where the camera is.
+    bool first_person_engaged = false;
     // Changes whenever the stable presentation-safety state changes. Ordinary
     // per-frame scene/camera publication does not advance it.
     uint64_t safety_generation = 1;
@@ -115,6 +124,13 @@ struct MkwVRPolicySnapshot {
     // the current presentation mode, so a transient scene/camera mismatch
     // cannot accept an immersive packet from an adjacent asynchronous frame.
     uint64_t content_tag = 0;
+
+    // The scale headset translation and IPD are converted at. Head offsets in
+    // metres must use the same value, or the camera and the world disagree.
+    float EffectiveUnitsPerMeter() const noexcept {
+        return first_person_engaged ? config.first_person_units_per_meter
+                                    : config.world_units_per_meter;
+    }
 };
 
 // All policy functions are thread-safe. Publishing functions are intended for
@@ -127,6 +143,14 @@ void MkwVRPolicySetAvailableBindings(uint32_t bindings) noexcept;
 void MkwVRPolicyPublishScene(const MkwVRSceneObservation& scene) noexcept;
 void MkwVRPolicyPublishRaceCamera(const MkwVRCameraObservation& camera) noexcept;
 void MkwVRPolicyInvalidateRaceCamera() noexcept;
+// Reported by the integration layer each time the first-person anchor engages
+// or disengages. It selects the world scale and nothing else: presentation
+// safety is unaffected, so this never advances the safety generation.
+void MkwVRPolicySetFirstPersonEngaged(bool engaged) noexcept;
+// Live world scale for the first-person camera. Separate from
+// MkwVRPolicyConfigure so the F10 slider can retune it during a race without
+// republishing (and revalidating) the whole configuration.
+void MkwVRPolicySetFirstPersonUnitsPerMeter(float units_per_meter) noexcept;
 MkwVRPolicySnapshot MkwVRPolicyGetSnapshot() noexcept;
 
 // This classifier is deliberately structural rather than heuristic: future

runtime/src/hle/vi.cpp

@@ -8,6 +8,7 @@
 #include "fiber_manager.h"
 #include "platform/host_platform.h"
 #include "runtime_log.h"
+#include "vr/mkw_vr_first_person.h"
 #include "vr/mkw_vr_policy.h"
 #include "vr/openxr_integration.h"
 
@@ -516,6 +517,27 @@ void PaceToRetraceBoundary(Clock::time_point deadline) {
     VI_HLE_ProcessRetracesDeferred(1);
 }
 
+// Hands Aurora the first-person camera relocation observed while this frame's
+// GX commands were produced. It has to be published here rather than by the XR
+// pacing thread: the anchor only makes sense against the recorded camera of
+// this exact frame, and the pacing thread does not know which frame its packet
+// will be paired with.
+void PublishVrSceneAnchor() {
+    static bool s_engaged = false;
+    const mkw::vr::FirstPersonAnchor anchor = mkw::vr::MkwVRFirstPersonGetAnchor();
+    aurora_set_stereo_scene_anchor(anchor.valid ? anchor.anchor_from_scene.data() : nullptr);
+
+    const bool engaged = anchor.valid;
+    if (engaged == s_engaged) {
+        return;
+    }
+    s_engaged = engaged;
+    // The world scale changes with the camera, and the virtual screen's metres
+    // are converted at that scale, so the two have to move together.
+    mkw::vr::MkwVRPolicySetFirstPersonEngaged(engaged);
+    settings_overlay::RefreshVrHudVirtualScreen();
+}
+
 } // namespace
 
 // Single owner of the Aurora frame presentation sequence: seals the active frame, optionally paces the
@@ -584,6 +606,7 @@ void VI_HLE_PresentFrame(bool presentedXfb, bool paceToRetrace) {
     }
 
     mkw::vr::OpenXRServiceProducerFrameBoundary();
+    PublishVrSceneAnchor();
     // Latch the current policy safety state into this exact Aurora job. The
     // asynchronous worker may ask for an XR packet after the guest has already
     // begun the next frame, so immersive replay is accepted only when both

runtime/src/settings_overlay.cpp

@@ -5,6 +5,8 @@
 #include "music_attenuation.h"
 #include "runtime_config.h"
 #include "runtime_log.h"
+#include "vr/mkw_vr_first_person.h"
+#include "vr/mkw_vr_policy.h"
 #include "wii_remote_input.h"
 
 #include <imgui.h>
@@ -103,6 +105,11 @@ bool g_vrEnabled = RuntimeConfigFile::VrEnabled(false);
 bool g_vrStopAtDisplayCopy = RuntimeConfigFile::VrStopAtDisplayCopy(true);
 bool g_vrSkipCopyClears = RuntimeConfigFile::VrSkipCopyClears(true);
 bool g_vrHudVirtualScreen = RuntimeConfigFile::VrHudVirtualScreen(true);
+bool g_vrFirstPerson = RuntimeConfigFile::VrFirstPerson(false);
+float g_vrFirstPersonUnitsPerMeter = RuntimeConfigFile::VrFirstPersonUnitsPerMeter(10.0f);
+float g_vrFirstPersonHeadUp = RuntimeConfigFile::VrFirstPersonHeadUpMeters(1.0f);
+float g_vrFirstPersonHeadForward = RuntimeConfigFile::VrFirstPersonHeadForwardMeters(0.0f);
+float g_vrFirstPersonHeadRight = RuntimeConfigFile::VrFirstPersonHeadRightMeters(0.0f);
 uint32_t g_disabledPostProcessingPaths = RuntimeConfigFile::DisabledPostProcessingPaths(0);
 std::array<int32_t, PAD_MAX_CONTROLLERS> g_configuredControllerIndices = [] {
     std::array<int32_t, PAD_MAX_CONTROLLERS> indices{};
@@ -708,9 +715,11 @@ void DrawAudioSettings() {
 
 // The virtual screen's placement comes from the launch-time [vr] geometry, the
 // same metres the menu quad is built from, converted into the world units the
-// eye replay works in.
+// eye replay works in. Those units follow the camera: the first-person view
+// renders at its own scale, and the screen has to be sized at the same one or
+// it would not stay 2 m across in front of the player.
 void ApplyVrHudVirtualScreen() {
-    const float unitsPerMeter = RuntimeConfigFile::VrWorldUnitsPerMeter(500.0f);
+    const float unitsPerMeter = mkw::vr::MkwVRPolicyGetSnapshot().EffectiveUnitsPerMeter();
     aurora_set_stereo_hud_screen(g_vrHudVirtualScreen,
                                  RuntimeConfigFile::VrHudWidthMeters(2.4f) * unitsPerMeter,
                                  RuntimeConfigFile::VrHudDistanceMeters(2.0f) * unitsPerMeter);
@@ -851,6 +860,51 @@ void DrawGraphicsSettings() {
             "the whole view. Its size and distance are the [vr] hud_width_meters and "
             "hud_distance_meters read at launch.");
     }
+    ImGui::Separator();
+    ImGui::Text("VR camera");
+    if (ImGui::Checkbox("First-person camera", &g_vrFirstPerson)) {
+        RuntimeConfigFile::SetVrFirstPerson(g_vrFirstPerson);
+        mkw::vr::MkwVRFirstPersonApplyConfiguredSettings();
+    }
+    if (ImGui::IsItemHovered()) {
+        ImGui::SetTooltip(
+            "Moves the camera to the Player 1 driver's head and keeps the horizon level, "
+            "instead of riding behind the kart. Applies during a single-screen race; menus "
+            "and split-screen are unaffected. The world scale below replaces "
+            "world_units_per_meter while it is engaged.");
+    }
+    // These are the tuning loop for the anchor: the right head height is a
+    // per-taste value that can only really be judged from inside the headset.
+    if (ImGui::SliderFloat("World units per metre (first person)", &g_vrFirstPersonUnitsPerMeter,
+                           1.0f, 200.0f, "%.1f")) {
+        RuntimeConfigFile::SetVrFirstPersonUnitsPerMeter(g_vrFirstPersonUnitsPerMeter);
+        mkw::vr::MkwVRFirstPersonApplyConfiguredSettings();
+        // The virtual screen's metres are converted at this same scale.
+        ApplyVrHudVirtualScreen();
+    }
+    if (ImGui::IsItemHovered()) {
+        ImGui::SetTooltip(
+            "Mario Kart Wii is authored at about 10 units per metre, which is what makes the "
+            "race read life-size. Raising this shrinks the world around you.");
+    }
+    bool headOffsetsChanged = false;
+    headOffsetsChanged |=
+        ImGui::SliderFloat("Head height (m)", &g_vrFirstPersonHeadUp, -1.0f, 3.0f, "%.2f");
+    headOffsetsChanged |=
+        ImGui::SliderFloat("Head forward (m)", &g_vrFirstPersonHeadForward, -20.0f, 20.0f, "%.2f");
+    headOffsetsChanged |=
+        ImGui::SliderFloat("Head sideways (m)", &g_vrFirstPersonHeadRight, -3.0f, 3.0f, "%.2f");
+    if (headOffsetsChanged) {
+        RuntimeConfigFile::SetVrFirstPersonHeadUpMeters(g_vrFirstPersonHeadUp);
+        RuntimeConfigFile::SetVrFirstPersonHeadForwardMeters(g_vrFirstPersonHeadForward);
+        RuntimeConfigFile::SetVrFirstPersonHeadRightMeters(g_vrFirstPersonHeadRight);
+        mkw::vr::MkwVRFirstPersonApplyConfiguredSettings();
+    }
+    ImGui::PushTextWrapPos(ImGui::GetCursorPosX() + 380.0f);
+    ImGui::TextDisabled(
+        "Where the head sits in the kart's own frame. Nudge it forward if the driver's own "
+        "head intrudes on the view.");
+    ImGui::PopTextWrapPos();
 }
 
 void DrawFpsOverlay() {
@@ -1072,11 +1126,14 @@ void InitializeRuntimeSettings() noexcept {
     aurora_set_stereo_skip_copy_clears(g_vrSkipCopyClears);
     ApplyVrHudVirtualScreen();
     aurora_set_skip_unready_pipelines(g_skipUnreadyPipelines);
+    mkw::vr::MkwVRFirstPersonApplyConfiguredSettings();
     g_strapInputAccepted.store(false, std::memory_order_relaxed);
     g_startupDismissFrame.store(UINT64_MAX, std::memory_order_relaxed);
     PADBlockInput(false);
 }
 
+void RefreshVrHudVirtualScreen() noexcept { ApplyVrHudVirtualScreen(); }
+
 void HandleEvents(const AuroraEvent* events) noexcept {
     if (!events) {
         return;

runtime/src/vr/mkw_vr_first_person.cpp

@@ -0,0 +1,298 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include "vr/mkw_vr_first_person.h"
+
+#include "memory.h"
+#include "runtime_config.h"
+#include "runtime_log.h"
+#include "vr/mkw_vr_policy.h"
+
+#include <mutex>
+
+extern "C" void func_805A6C58(CpuContext* context);
+
+namespace mkw::vr {
+namespace {
+
+// ---------------------------------------------------------------------------
+// PAL RMCP01 object layout.
+//
+// Derived from the shipped StaticR.rel and cross-checked against the mkw
+// decompilation. Each constant names the accessor that proves it, so a future
+// region or a mod that moves these can be re-derived the same way. Keep in
+// sync with projects/mkwii/MAP.txt and the generated translations.
+// ---------------------------------------------------------------------------
+
+// RaceCamera::GetViewMtx (0x805A6C58) writes the authoritative view matrix to
+// its r4 output buffer. The adjacent RaceCamera fields are state vectors, not
+// a view matrix, so call the game's getter instead of guessing an object offset.
+constexpr uint32_t kRaceCameraScratchBytes = 0x300u;
+
+// Kart::Manager's instance pointer. Its CreateInstance (0x8058FAA8) resolves
+// the slot as 0x809C0000 + 6392 in the generated translation. Read directly
+// rather than observed from Kart::Manager::Update's r3, so enabling the camera
+// needs no change to the translated output: an entry observer only exists in a
+// build whose translation was regenerated for it, and its absence is silent.
+// This mirrors how the race scene's instance slot is reached in
+// mkw_vr_instrumentation.cpp.
+constexpr uint32_t kKartManagerInstanceAddress = 0x809C18F8u;
+// Kart::Manager::GetKartPlayer (0x80590100): `lwz r3,0x20(r3)` then indexes.
+constexpr uint32_t kKartManagerPlayersOffset = 0x20u;
+// Kart::Link::GetKartPosition (0x8059020C) walks proxy -> accessor -> body ->
+// physics -> dynamics; the first three links are shared by every kart accessor.
+constexpr uint32_t kKartProxyAccessorOffset = 0x00u;
+constexpr uint32_t kKartAccessorBodyOffset = 0x08u;
+constexpr uint32_t kKartBodyPhysicsOffset = 0x90u;
+// KartPhysics::pose (Kart::Link::GetMtx 0x80590264). This is the physics-driven
+// pose, deliberately not the visual one: an animated frame would bob the
+// camera. Kart::Link::GetKartBodyMtx (0x80590278) returns KartBody+0x1C, the
+// visual pose, and is the alternative to try if the seat ever looks detached.
+constexpr uint32_t kKartPhysicsPoseOffset = 0x9Cu;
+
+// Offline Mario Kart Wii puts the local racer first, and immersive
+// presentation already requires exactly one on-screen player.
+constexpr uint32_t kLocalPlayerIndex = 0;
+
+// Frames the last good anchor survives a failed read before the camera returns
+// to the game's own. Rides out a transient null during a respawn or transition
+// without letting a genuinely broken anchor persist.
+constexpr int kHoldFrames = 10;
+
+// ---------------------------------------------------------------------------
+// Guest reads. Everything is bounds-checked and exception-guarded so a pointer
+// caught mid-teardown can only cost this frame's anchor.
+// ---------------------------------------------------------------------------
+
+bool ReadGuestPointer(uint32_t address, uint32_t& out) noexcept {
+    return Memory::TryRead32(address, out) && out != 0;
+}
+
+constexpr uint32_t kMtx34Bytes = 12u * sizeof(float);
+
+bool ReadGuestMtx34(uint32_t address, Mtx34& out) noexcept {
+    if (address == 0 || !Memory::Contains(address, kMtx34Bytes)) {
+        return false;
+    }
+    try {
+        for (uint32_t i = 0; i < out.size(); ++i) {
+            out[i] = Memory::ReadFloat32(address + i * static_cast<uint32_t>(sizeof(float)));
+        }
+    } catch (const Memory::AccessViolation&) {
+        return false;
+    }
+    return detail::IsFiniteMtx34(out);
+}
+
+bool ReadRaceCameraViewMatrix(const CpuContext* context, uint32_t camera_address,
+                              Mtx34& out) noexcept {
+    if (context == nullptr || camera_address == 0 ||
+        context->gpr[1] < kRaceCameraScratchBytes) {
+        return false;
+    }
+
+    CpuContext call_context = *context;
+    const uint32_t scratch = context->gpr[1] - kRaceCameraScratchBytes;
+    call_context.gpr[3] = camera_address;
+    call_context.gpr[4] = scratch;
+    call_context.gpr[5] = scratch + 48u;
+    try {
+        CpuContextScope scope(&call_context);
+        func_805A6C58(&call_context);
+        return ReadGuestMtx34(scratch, out);
+    } catch (const Memory::AccessViolation&) {
+        return false;
+    }
+}
+
+// The pointer walk, kept inspectable: on failure `failed_step` names the link
+// that broke and the resolved pointers before it are still filled in. One log
+// line then says exactly which offset needs revisiting.
+struct KartPoseRead {
+    const char* failed_step = nullptr;
+    uint32_t manager = 0;
+    uint32_t players = 0;
+    uint32_t proxy = 0;
+    uint32_t accessor = 0;
+    uint32_t body = 0;
+    uint32_t physics = 0;
+};
+
+KartPoseRead ReadPlayerKartPose(Mtx34& out) noexcept {
+    KartPoseRead read{};
+    if (!ReadGuestPointer(kKartManagerInstanceAddress, read.manager)) {
+        read.failed_step = "Kart::Manager instance";
+    } else if (!ReadGuestPointer(read.manager + kKartManagerPlayersOffset, read.players)) {
+        read.failed_step = "Kart::Manager players array";
+    } else if (!ReadGuestPointer(read.players + kLocalPlayerIndex * 4u, read.proxy)) {
+        read.failed_step = "player kart object";
+    } else if (!ReadGuestPointer(read.proxy + kKartProxyAccessorOffset, read.accessor)) {
+        read.failed_step = "kart accessor";
+    } else if (!ReadGuestPointer(read.accessor + kKartAccessorBodyOffset, read.body)) {
+        read.failed_step = "kart body";
+    } else if (!ReadGuestPointer(read.body + kKartBodyPhysicsOffset, read.physics)) {
+        read.failed_step = "kart physics";
+    } else if (!ReadGuestMtx34(read.physics + kKartPhysicsPoseOffset, out)) {
+        read.failed_step = "kart pose matrix";
+    }
+    return read;
+}
+
+// ---------------------------------------------------------------------------
+
+struct FirstPersonState {
+    bool enabled = false;
+    FirstPersonHeadOffsets offsets{};
+    float units_per_meter = 10.0f;
+
+    uint32_t camera_address = 0;
+
+    FirstPersonAnchor anchor{};
+    int hold_frames = 0;
+    bool ever_valid_this_race = false;
+    bool failure_logged = false;
+    uint64_t logged_frame = 0;
+};
+
+std::mutex g_mutex;
+FirstPersonState g_state;
+
+void LogAnchorLocked(uint64_t frame, const Mtx34& anchor, const Mtx34& view_from_world,
+                     const KartPoseRead& kart, const Mtx34& kart_from_local) noexcept {
+    // One line per second at 60 Hz: enough to confirm the offsets on-device
+    // without drowning the log during a race.
+    if (g_state.logged_frame != 0 && frame - g_state.logged_frame < 60) {
+        return;
+    }
+    g_state.logged_frame = frame;
+    // The anchor's translation is -R*a, so negating it gives the head's offset
+    // from the recorded camera measured in the levelled camera's own axes.
+    // While driving it should stay roughly constant: a little to the side, a
+    // little below the chase camera, and well in front of it.
+    RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] first-person anchor: frame=" << frame << ", camera=0x"
+                           << std::hex << g_state.camera_address << std::dec
+                           << ", head from camera (right, up, forward)=(" << -anchor[3] << ", "
+                           << -anchor[7] << ", " << anchor[11] << ") units" << std::endl;
+    RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] first-person view: rows=(" << view_from_world[0] << ", "
+                           << view_from_world[1] << ", " << view_from_world[2] << "; "
+                           << view_from_world[4] << ", " << view_from_world[5] << ", "
+                           << view_from_world[6] << "; " << view_from_world[8] << ", "
+                           << view_from_world[9] << ", " << view_from_world[10]
+                           << "), translation=(" << view_from_world[3] << ", "
+                           << view_from_world[7] << ", " << view_from_world[11] << ")"
+                           << std::endl;
+    RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] first-person pose: physics=0x" << std::hex << kart.physics
+                           << ", pose=0x" << (kart.physics + kKartPhysicsPoseOffset) << std::dec
+                           << ", rows=(" << kart_from_local[0] << ", " << kart_from_local[1]
+                           << ", " << kart_from_local[2] << "; " << kart_from_local[4] << ", "
+                           << kart_from_local[5] << ", " << kart_from_local[6] << "; "
+                           << kart_from_local[8] << ", " << kart_from_local[9] << ", "
+                           << kart_from_local[10] << "), translation=(" << kart_from_local[3]
+                           << ", " << kart_from_local[7] << ", " << kart_from_local[11] << ")"
+                           << std::endl;
+    RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] first-person pose bits: translation=(0x"
+                           << std::hex << std::bit_cast<uint32_t>(kart_from_local[3]) << ", 0x"
+                           << std::bit_cast<uint32_t>(kart_from_local[7]) << ", 0x"
+                           << std::bit_cast<uint32_t>(kart_from_local[11]) << ")" << std::dec
+                           << std::endl;
+}
+
+} // namespace
+
+void MkwVRFirstPersonConfigure(bool enabled, const FirstPersonHeadOffsets& offsets,
+                               float units_per_meter) noexcept {
+    std::lock_guard lock(g_mutex);
+    g_state.enabled = enabled;
+    g_state.offsets = offsets;
+    if (detail::IsFiniteFloat(&units_per_meter) && units_per_meter > 0.0f) {
+        g_state.units_per_meter = units_per_meter;
+    }
+    if (!enabled) {
+        g_state.anchor = {};
+        g_state.hold_frames = 0;
+    }
+}
+
+void MkwVRFirstPersonApplyConfiguredSettings() noexcept {
+    const float units_per_meter = RuntimeConfigFile::VrFirstPersonUnitsPerMeter(10.0f);
+    const FirstPersonHeadOffsets offsets{
+        RuntimeConfigFile::VrFirstPersonHeadRightMeters(0.0f),
+        RuntimeConfigFile::VrFirstPersonHeadUpMeters(1.0f),
+        RuntimeConfigFile::VrFirstPersonHeadForwardMeters(0.0f),
+    };
+    MkwVRFirstPersonConfigure(RuntimeConfigFile::VrFirstPerson(false), offsets, units_per_meter);
+    MkwVRPolicySetFirstPersonUnitsPerMeter(units_per_meter);
+}
+
+void MkwVRFirstPersonReset() noexcept {
+    std::lock_guard lock(g_mutex);
+    g_state.camera_address = 0;
+    g_state.anchor = {};
+    g_state.hold_frames = 0;
+    g_state.ever_valid_this_race = false;
+    g_state.failure_logged = false;
+    g_state.logged_frame = 0;
+}
+
+void MkwVRFirstPersonUpdate(uint64_t guest_frame_index, uint32_t race_camera_address) noexcept {
+    std::lock_guard lock(g_mutex);
+    if (!g_state.enabled) {
+        g_state.anchor = {};
+        g_state.hold_frames = 0;
+        return;
+    }
+    g_state.camera_address = race_camera_address;
+
+    Mtx34 view_from_world{};
+    Mtx34 kart_from_local{};
+    Mtx34 anchor{};
+    KartPoseRead kart{};
+    const char* failed_step = nullptr;
+    if (race_camera_address == 0) {
+        failed_step = "race camera (none updated this frame)";
+    } else if (!ReadRaceCameraViewMatrix(TryGetCpuContext(), race_camera_address,
+                                         view_from_world)) {
+        failed_step = "race camera view matrix";
+    } else if (kart = ReadPlayerKartPose(kart_from_local); kart.failed_step != nullptr) {
+        failed_step = kart.failed_step;
+    } else if (!ComputeFirstPersonAnchor(view_from_world, kart_from_local,
+                                         g_state.offsets.right * g_state.units_per_meter,
+                                         g_state.offsets.up * g_state.units_per_meter,
+                                         g_state.offsets.forward * g_state.units_per_meter,
+                                         /*level_horizon=*/true, anchor)) {
+        failed_step = "anchor math (degenerate camera or kart frame)";
+    }
+
+    if (failed_step == nullptr) {
+        g_state.anchor = {anchor, true, guest_frame_index};
+        g_state.hold_frames = kHoldFrames;
+        g_state.ever_valid_this_race = true;
+        LogAnchorLocked(guest_frame_index, anchor, view_from_world, kart, kart_from_local);
+        return;
+    }
+
+    if (g_state.hold_frames > 0) {
+        --g_state.hold_frames;
+        g_state.anchor.guest_frame_index = guest_frame_index;
+        return;
+    }
+    if (!g_state.ever_valid_this_race && !g_state.failure_logged) {
+        // Once per race, naming the exact link that broke: every address below
+        // is a PAL RMCP01 constant, so this is what says which one to revisit.
+        g_state.failure_logged = true;
+        RT_LOG(RT_TAG_RUNTIME)
+            << "[mkw-vr] first-person camera is enabled but could not resolve the "
+            << failed_step << "; staying on the game's own camera (camera=0x" << std::hex
+            << race_camera_address << ", manager=0x" << kart.manager << ", players=0x"
+            << kart.players << ", kart=0x" << kart.proxy << ", accessor=0x" << kart.accessor
+            << ", body=0x" << kart.body << ", physics=0x" << kart.physics << std::dec << ")"
+            << std::endl;
+    }
+    g_state.anchor = {};
+}
+
+FirstPersonAnchor MkwVRFirstPersonGetAnchor() noexcept {
+    std::lock_guard lock(g_mutex);
+    return g_state.anchor;
+}
+
+} // namespace mkw::vr

runtime/src/vr/mkw_vr_instrumentation.cpp

@@ -5,6 +5,7 @@
 #include "memory.h"
 #include "ppc_runtime.h"
 #include "runtime_log.h"
+#include "vr/mkw_vr_first_person.h"
 #include "vr/mkw_vr_policy.h"
 
 #include <algorithm>
@@ -69,6 +70,15 @@ uint32_t CameraCount(uint64_t frame) noexcept {
     return g_instrumentation.camera_count;
 }
 
+// The first RaceCamera updated this frame. The game also updates cameras for
+// transitions and effects, so first-wins is what keeps the first-person anchor
+// deterministic: Mario Kart updates the racers' cameras before those.
+uint32_t FirstCamera(uint64_t frame) noexcept {
+    std::lock_guard lock(g_instrumentation_mutex);
+    ResetCamerasForFrameLocked(frame);
+    return g_instrumentation.camera_count != 0 ? g_instrumentation.cameras[0] : 0;
+}
+
 uint32_t RaceScreenCount() noexcept {
     uint32_t race_scene = 0;
     if (!Memory::TryRead32(kRaceSceneInstanceAddress, race_scene) || race_scene == 0 ||
@@ -150,6 +160,7 @@ extern "C" void MkwVRObserveTranslatedFunctionEntry(uint32_t address,
         }
         PublishRaceScene(frame, 0);
         MkwVRPolicyInvalidateRaceCamera();
+        MkwVRFirstPersonReset();
         RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] entered RaceScene at frame " << frame
                                << std::endl;
         break;
@@ -169,6 +180,10 @@ extern "C" void MkwVRObserveTranslatedFunctionEntry(uint32_t address,
         const uint32_t screen_count = RaceScreenCount();
         LogRaceEvidenceIfChanged(frame, screen_count, camera_count);
         PublishRaceScene(frame, screen_count);
+        // Every kart and camera has been updated for this frame and none of
+        // the frame's draws have been issued yet, so the values behind these
+        // pointers are exactly the ones those draws will use.
+        MkwVRFirstPersonUpdate(frame, FirstCamera(frame));
         break;
     }
     case kRaceSceneOnExit: {
@@ -177,6 +192,7 @@ extern "C" void MkwVRObserveTranslatedFunctionEntry(uint32_t address,
         scene.guest_frame_index = frame;
         MkwVRPolicyPublishScene(scene);
         MkwVRPolicyInvalidateRaceCamera();
+        MkwVRFirstPersonReset();
         RT_LOG(RT_TAG_RUNTIME) << "[mkw-vr] exited RaceScene at frame " << frame
                                << std::endl;
         break;

runtime/src/vr/mkw_vr_policy.cpp

@@ -17,6 +17,7 @@ struct PolicyState {
     MkwVRCameraObservation camera{};
     uint32_t available_bindings = MkwVRBindingNone;
     bool session_active = false;
+    bool first_person_engaged = false;
     uint64_t safety_generation = 1;
 };
 
@@ -56,6 +57,9 @@ MkwVRPolicyConfig SanitizeConfig(const MkwVRPolicyConfig& config) noexcept {
     if (!IsFinitePositive(&sanitized.hud_scale)) {
         sanitized.hud_scale = kDefaultConfig.hud_scale;
     }
+    if (!IsFinitePositive(&sanitized.first_person_units_per_meter)) {
+        sanitized.first_person_units_per_meter = kDefaultConfig.first_person_units_per_meter;
+    }
     return sanitized;
 }
 
@@ -231,6 +235,21 @@ void MkwVRPolicyInvalidateRaceCamera() noexcept {
     ApplyPolicyMutation([&] { g_policy.camera.valid = false; });
 }
 
+void MkwVRPolicySetFirstPersonEngaged(bool engaged) noexcept {
+    std::lock_guard<std::mutex> lock(g_policy_mutex);
+    // Not routed through ApplyPolicyMutation: where the camera sits does not
+    // change which content is safe to present, and advancing the safety
+    // generation here would drop a frame to mono on every engage.
+    g_policy.first_person_engaged = engaged;
+}
+
+void MkwVRPolicySetFirstPersonUnitsPerMeter(float units_per_meter) noexcept {
+    std::lock_guard<std::mutex> lock(g_policy_mutex);
+    if (IsFinitePositive(&units_per_meter)) {
+        g_policy.config.first_person_units_per_meter = units_per_meter;
+    }
+}
+
 MkwVRPolicySnapshot MkwVRPolicyGetSnapshot() noexcept {
     std::lock_guard<std::mutex> lock(g_policy_mutex);
     MkwVRPolicySnapshot snapshot;
@@ -240,6 +259,8 @@ MkwVRPolicySnapshot MkwVRPolicyGetSnapshot() noexcept {
     snapshot.camera = g_policy.camera;
     snapshot.available_bindings = g_policy.available_bindings;
     snapshot.session_active = g_policy.session_active;
+    snapshot.first_person_engaged =
+        g_policy.first_person_engaged && snapshot.presentation == VRPresentationMode::ImmersiveRace;
     snapshot.safety_generation = g_policy.safety_generation;
     snapshot.content_tag = MakeContentTag(g_policy, snapshot.presentation);
     return snapshot;

runtime/src/vr/openxr_integration.cpp

@@ -8,6 +8,7 @@
 
 #include "runtime_config.h"
 #include "runtime_log.h"
+#include "vr/mkw_vr_first_person.h"
 #include "vr/mkw_vr_policy.h"
 #include "vr/mkw_vr_instrumentation.h"
 
@@ -44,8 +45,10 @@ void ConfigurePolicy(bool enabled) noexcept {
     config.world_units_per_meter = RuntimeConfigFile::VrWorldUnitsPerMeter(500.0f);
     config.hud_distance_meters = RuntimeConfigFile::VrHudDistanceMeters(2.0f);
     config.hud_width_meters = RuntimeConfigFile::VrHudWidthMeters(2.4f);
+    config.first_person_units_per_meter = RuntimeConfigFile::VrFirstPersonUnitsPerMeter(10.0f);
     MkwVRPolicyConfigure(config);
     MkwVRInstrumentationInitialize();
+    MkwVRFirstPersonApplyConfiguredSettings();
 }
 
 #if defined(MKW_ENABLE_OPENXR) && defined(_WIN32)
@@ -487,7 +490,11 @@ private:
 
             {
                 std::lock_guard lock(published_mutex_);
-                BuildPublishedFrame(frame, immersive, policy.config.world_units_per_meter,
+                // First person renders at life-size scale, third person at the
+                // configured diorama scale. Head translation and IPD are the
+                // only things this multiplies, so a one-frame disagreement with
+                // the camera's own switch is not observable.
+                BuildPublishedFrame(frame, immersive, policy.EffectiveUnitsPerMeter(),
                                     policy.content_tag);
                 published_.store(&published_frame_, std::memory_order_release);
             }

runtime/tests/vr_first_person_tests.cpp

@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+//
+// The first-person VR camera's transform, tested without a guest. Everything
+// here exercises ComputeFirstPersonAnchor, which turns the game's own view and
+// kart matrices into the relocation Aurora composes onto each eye.
+
+#include "vr/mkw_vr_first_person.h"
+
+#include <cmath>
+#include <initializer_list>
+#include <iostream>
+#include <limits>
+
+namespace {
+
+using mkw::vr::ComputeFirstPersonAnchor;
+using mkw::vr::kIdentityMtx34;
+using mkw::vr::Mtx34;
+
+int g_failures = 0;
+
+void Check(bool condition, const char* what) {
+    if (!condition) {
+        ++g_failures;
+        std::cerr << "FAILED: " << what << '\n';
+    }
+}
+
+void CheckNear(float actual, float expected, const char* what, float tolerance = 1.0e-3f) {
+    if (!(std::fabs(actual - expected) <= tolerance)) {
+        ++g_failures;
+        std::cerr << "FAILED: " << what << " (expected " << expected << ", got " << actual << ")\n";
+    }
+}
+
+// out = matrix * (x, y, z, 1)
+void Apply(const Mtx34& matrix, float x, float y, float z, float out[3]) {
+    out[0] = matrix[0] * x + matrix[1] * y + matrix[2] * z + matrix[3];
+    out[1] = matrix[4] * x + matrix[5] * y + matrix[6] * z + matrix[7];
+    out[2] = matrix[8] * x + matrix[9] * y + matrix[10] * z + matrix[11];
+}
+
+// A view matrix for a camera at `eye` looking along -Z with no pitch or roll.
+Mtx34 LevelViewAt(float x, float y, float z) {
+    Mtx34 view = kIdentityMtx34;
+    view[3] = -x;
+    view[7] = -y;
+    view[11] = -z;
+    return view;
+}
+
+// The same, pitched down by `radians` about the view's X axis. Rows are the
+// camera's axes in world space, which is what a world -> view matrix holds.
+Mtx34 PitchedViewAt(float x, float y, float z, float radians) {
+    const float c = std::cos(radians);
+    const float s = std::sin(radians);
+    Mtx34 view{};
+    view[0] = 1.0f;
+    view[5] = c;
+    view[6] = s;
+    view[9] = -s;
+    view[10] = c;
+    view[3] = -(view[0] * x + view[1] * y + view[2] * z);
+    view[7] = -(view[4] * x + view[5] * y + view[6] * z);
+    view[11] = -(view[8] * x + view[9] * y + view[10] * z);
+    return view;
+}
+
+Mtx34 KartAt(float x, float y, float z) {
+    Mtx34 pose = kIdentityMtx34;
+    pose[3] = x;
+    pose[7] = y;
+    pose[11] = z;
+    return pose;
+}
+
+void TestNeutralInputsProduceIdentity() {
+    Mtx34 anchor{};
+    Check(ComputeFirstPersonAnchor(kIdentityMtx34, kIdentityMtx34, 0.0f, 0.0f, 0.0f,
+                                   /*level_horizon=*/true, anchor),
+          "a camera already at the head must produce an anchor");
+    for (size_t i = 0; i < anchor.size(); ++i) {
+        CheckNear(anchor[i], kIdentityMtx34[i], "neutral inputs must produce the identity anchor");
+    }
+}
+
+void TestUnlevelledAnchorIsPureTranslation() {
+    // Camera 5 m behind and 2 m above the origin, kart at the origin, head 1 m up.
+    const Mtx34 view = LevelViewAt(0.0f, 2.0f, 5.0f);
+    const Mtx34 kart = KartAt(0.0f, 0.0f, 0.0f);
+    Mtx34 anchor{};
+    Check(ComputeFirstPersonAnchor(view, kart, 0.0f, 1.0f, 0.0f, /*level_horizon=*/false, anchor),
+          "an unlevelled anchor must be computable");
+
+    // The head sits at (0, -1, -5) in view space, so the anchor's translation
+    // is its negation.
+    CheckNear(anchor[3], 0.0f, "no lateral offset");
+    CheckNear(anchor[7], 1.0f, "the anchor cancels the head's -1 view-space height");
+    CheckNear(anchor[11], 5.0f, "the anchor cancels the head's -5 view-space depth");
+
+    // Rotation untouched, so a world point keeps its orientation and only shifts.
+    float moved[3];
+    Apply(anchor, 0.0f, -1.0f, -5.0f, moved);
+    CheckNear(moved[0], 0.0f, "the head lands at the eye origin (x)");
+    CheckNear(moved[1], 0.0f, "the head lands at the eye origin (y)");
+    CheckNear(moved[2], 0.0f, "the head lands at the eye origin (z)");
+}
+
+void TestLevellingRemovesCameraPitch() {
+    // A chase camera looking down at the kart, which is the ordinary Mario Kart
+    // Wii case: first person must not inherit that downward tilt.
+    const float pitch = 0.35f;
+    const Mtx34 view = PitchedViewAt(0.0f, 2.0f, 5.0f, pitch);
+    const Mtx34 kart = KartAt(0.0f, 0.0f, 0.0f);
+    Mtx34 anchor{};
+    Check(ComputeFirstPersonAnchor(view, kart, 0.0f, 1.0f, 0.0f, /*level_horizon=*/true, anchor),
+          "a pitched camera must still produce an anchor");
+
+    // The anchored camera's axes, expressed in world space: rows of A_rot times
+    // the view rotation. Its forward is -row2, and it must be horizontal.
+    const float worldUp[3]{0.0f, 1.0f, 0.0f};
+    float rowInWorld[3][3];
+    for (size_t row = 0; row < 3; ++row) {
+        for (size_t axis = 0; axis < 3; ++axis) {
+            // view's rows are the camera axes in world space, so a view-space
+            // vector returns to world space through view's transpose.
+            rowInWorld[row][axis] = anchor[row * 4 + 0] * view[0 * 4 + axis] +
+                                    anchor[row * 4 + 1] * view[1 * 4 + axis] +
+                                    anchor[row * 4 + 2] * view[2 * 4 + axis];
+        }
+    }
+    const float forwardDotUp = -(rowInWorld[2][0] * worldUp[0] + rowInWorld[2][1] * worldUp[1] +
+                                 rowInWorld[2][2] * worldUp[2]);
+    CheckNear(forwardDotUp, 0.0f, "the levelled forward axis must be horizontal");
+    const float rightDotUp = rowInWorld[0][0] * worldUp[0] + rowInWorld[0][1] * worldUp[1] +
+                             rowInWorld[0][2] * worldUp[2];
+    CheckNear(rightDotUp, 0.0f, "the levelled right axis must be horizontal");
+    const float upDotUp = rowInWorld[1][0] * worldUp[0] + rowInWorld[1][1] * worldUp[1] +
+                          rowInWorld[1][2] * worldUp[2];
+    CheckNear(upDotUp, 1.0f, "the levelled up axis must be world up");
+
+    // The head still lands exactly at the eye origin.
+    float head[3];
+    Apply(view, 0.0f, 1.0f, 0.0f, head);
+    float moved[3];
+    Apply(anchor, head[0], head[1], head[2], moved);
+    CheckNear(moved[0], 0.0f, "the head lands at the eye origin under levelling (x)");
+    CheckNear(moved[1], 0.0f, "the head lands at the eye origin under levelling (y)");
+    CheckNear(moved[2], 0.0f, "the head lands at the eye origin under levelling (z)");
+}
+
+void TestAnchorRotationStaysOrthonormal() {
+    // Straight down at the kart: the camera's own forward projects to nothing on
+    // the horizon plane, so the heading has to be recovered from its up axis.
+    const float kHalfPi = 1.57079632679f;
+    for (const float pitch : {0.0f, 0.35f, kHalfPi, -kHalfPi, 3.0f}) {
+        const Mtx34 view = PitchedViewAt(3.0f, 12.0f, -7.0f, pitch);
+        Mtx34 anchor{};
+        Check(ComputeFirstPersonAnchor(view, KartAt(3.0f, 0.0f, -20.0f), 0.1f, 1.0f, 0.2f,
+                                       /*level_horizon=*/true, anchor),
+              "every camera pitch must produce an anchor");
+        for (size_t row = 0; row < 3; ++row) {
+            for (size_t other = row; other < 3; ++other) {
+                float dot = 0.0f;
+                for (size_t axis = 0; axis < 3; ++axis) {
+                    dot += anchor[row * 4 + axis] * anchor[other * 4 + axis];
+                }
+                CheckNear(dot, row == other ? 1.0f : 0.0f,
+                          "the anchor's rotation must stay orthonormal");
+            }
+        }
+    }
+}
+
+void TestNonFiniteInputIsRejected() {
+    Mtx34 broken = kIdentityMtx34;
+    broken[3] = std::numeric_limits<float>::infinity();
+    Mtx34 anchor = kIdentityMtx34;
+    anchor[3] = 1234.0f;
+    Check(!ComputeFirstPersonAnchor(broken, kIdentityMtx34, 0.0f, 1.0f, 0.0f, true, anchor),
+          "a non-finite view matrix must be rejected");
+    CheckNear(anchor[3], 1234.0f, "a rejected anchor must leave the output untouched");
+}
+
+void TestDegenerateKartPoseIsRejected() {
+    Mtx34 collapsed{};
+    Mtx34 anchor{};
+    // A zeroed view matrix has no world up to level against.
+    Check(!ComputeFirstPersonAnchor(collapsed, kIdentityMtx34, 0.0f, 1.0f, 0.0f, true, anchor),
+          "a collapsed view matrix must be rejected");
+}
+
+} // namespace
+
+int main() {
+    TestNeutralInputsProduceIdentity();
+    TestUnlevelledAnchorIsPureTranslation();
+    TestLevellingRemovesCameraPitch();
+    TestAnchorRotationStaysOrthonormal();
+    TestNonFiniteInputIsRejected();
+    TestDegenerateKartPoseIsRejected();
+    if (g_failures != 0) {
+        std::cerr << g_failures << " check(s) failed\n";
+        return 1;
+    }
+    return 0;
+}