Adds quaternion-orientation plumbing for the Steam Controller and the
supporting infrastructure (Test Report multi-interface fan-out, runtime
axis-transform switcher). Pitch + roll track physical motion correctly
in the on-screen model when using the field-tested 'yzp-xy-alt'
transform. Yaw signal is captured but ~10× weaker than pitch/roll
(typical for IMUs without a magnetometer), and visually not convincing
— left as a known limitation for future iteration.

Driver (packages/core/src/drivers/steam-controller-driver.js):
- Parses 4× int16 LE at bytes 31-38 as quaternion components, returns
  parsed.orientation = {x, y, z, w}. The first component (bytes 31-32)
  is actually the high half of the device's uint32 timestamp on this
  firmware — SteamlessController's docs claim the quaternion lives at
  data[31-38] but variance analysis of a real Test Report capture
  shows bytes 29-32 are a ~1MHz uint32 timestamp. Three.js's normalize
  flattens the timestamp byte's effect enough that pitch + roll
  visualize correctly. Better encoding TBD — possibly Euler angles in
  radians, didn't lead to working motion when tried.
- emitsRawGyro = false (already) flag tells app.js to skip the
  rate-based calibration UX since the driver emits orientation, not
  rates.

App (apps/overlay/src/js/app.js):
- Quaternion fast-path: when parsed.orientation is present, writes
  directly into gyroFusion.orientation (bypasses SensorFusion). All
  downstream consumers (overlay body rotation, gimbal widget, R/P/Y
  readout, gyro HUD) pick it up unchanged.
- Reference-quaternion capture: first valid parsed.orientation
  becomes the rest reference; every subsequent frame is emitted as
  delta = current * ref⁻¹. L3+R3 (existing calibrate combo)
  recaptures the reference. Hooks into existing showCalibHint UX so
  user sees feedback when they recenter.
- Runtime axis-transform switcher: window.setSteamQuatTransform(mode)
  cycles between 10+ candidate IMU-body→visualizer-world basis
  transforms. Default 'yzp-xy-alt' = (z, x, y), found via empirical
  iteration. Other transforms in the registry for future tuning.

Test Report wizard (apps/overlay/src/js/test-report.js):
- Multi-interface input fan-out: previously the wizard listened on
  one Puck HID handle (whichever the picker returned) — that's the
  wrong interface for the Puck, so every step recorded 0 reports.
  Now mirrors the gyro pipeline's fan-out: enumerates approved
  siblings with same vid:pid, opens + attaches the recorder to each.
  Single-interface pads (DualSense filtered by usagePage) are no-op.

UI:
- Axis readout (P/R/Y degrees) moved from top-right to bottom-center.
  Was overlapping the Puck status banner.
- HUD Position setting renamed → "Roll HUD Position" since it only
  controls the gyro/roll arc widget, not the axis readout.

Verified via Test Report variance analysis on real hardware:
- Pitch movement primarily varies bytes 35-36 (353× baseline stddev)
- Roll movement primarily varies bytes 33-34 (467× baseline)
- Yaw movement varies all components weakly (max 59× baseline, but
  absolute stddev 10× smaller than pitch/roll)

Phase 3a is sized as "best-with-current-understanding" — the right
IMU encoding remains unclear and the test infrastructure to iterate
further is in place for a future contributor.