Fix pan axis drift: soft-iron bug + gyro calibration improvements

[sorcerdu]

Problem

On Seeed XIAO BLE Sense, the pan (yaw) axis exhibits ~0.1°/s cumulative drift even when the device is completely stationary. While the center-reset button can temporarily correct it, the continuous drift significantly degrades the user experience during extended sessions.

Root Cause Analysis

Two independent sources of drift were identified. Controlled testing on Seeed XIAO BLE Sense (stationary, bench-mounted):

Fix applied	Pan drift	Improvement
None (original)	~0.1°/s (6°/min)	—
Soft-iron fix only	~0.005°/s (0.1° per 20s, ~0.3°/min)	20×
Soft-iron + gyro improvements	<1° per hour (~0.017°/min)	360×

Bug 1: Soft-iron matrix in-place overwrite

The 3×3 soft-iron compensation matrix was applied in-place:

magx = (magx * magsioff[0]) + (magy * magsioff[1]) + (magz * magsioff[2]);
magy = (magx * magsioff[3]) + (magy * magsioff[4]) + (magz * magsioff[5]);  // uses updated magx!
magz = (magx * magsioff[6]) + (magy * magsioff[7]) + (magz * magsioff[8]);  // uses updated magx, magy!

Each row used already-overwritten values from previous rows. Even at rest, magnetometer noise combined with this wrong transform produced a fluctuating heading reference that the Madgwick filter integrated into slow yaw drift.

Bug 2: Noisy gyro bias estimate

The EMA-based bias estimator (alpha=0.05) effectively only averaged the last ~20 samples, leaving significant noise in the zero-bias estimate. Any residual error in the bias integrates directly into yaw over time.

Changes

• fix(sensor): Use temporary variables for soft-iron matrix multiplication (commit 1)
• feat(sensor): (commit 2)
  • Replace EMA with arithmetic mean over 1000 samples (~31× noise reduction on bias estimate)
  • Add automatic re-calibration when device is stationary for >5s (rate-limited to once per 5 minutes) to compensate for thermal drift
  • Add manual re-calibration on center-reset button press
  • Clamp Madgwick delta-t to handle thread scheduling delays and 32-bit timer wraparound

Testing Scope

Verified on Seeed XIAO BLE Sense only. The changes are board-agnostic (no platform-specific code modified). Testing on other supported boards (Arduino Nano 33 BLE, ESP32-C3 DevkitM, M5StickC Plus, RPi Pico W, DTQSys HT, Licardo HT) is welcomed from the community.

Notes

• GYRO_STABLE_SAMPLES increased from 400 to 1000, adding ~4s to boot calibration time (now ~6.7s @ 150Hz)
• Auto-recalibration only triggers when the device has been physically stationary for >5s; it will not interrupt active headtracking use
• Auto-recalibration uses raw sensor values with a generous threshold (5 dps) as a pre-filter; the precise stability check remains in gyroCalibrate() using board-specific GYRO_STABLE_DIFF
• No changes to the communication protocol, configuration interface, or saved flash format