Update impl spec with physics-based data generation and validation

Major changes to the data generator design:

- DrivingProfile now parameterizes TRAJECTORY (speed, curvature,
  smoothness, braking_intensity) instead of raw sensor values.
  Sensor readings are DERIVED using physics equations, making data
  authentic by construction.

- Added SENSOR_RANGES constants from real RC car hardware
  (IMU_GYRO_NORM=250, IMU_ACCEL_NORM=20, MAX_SPEED=4.4).

- Added validate_generated_tub() with 6 physics checks: sensor
  range bounds, speed continuity, distance monotonicity,
  gyro-curvature correlation, accel_x vs speed change correlation,
  and timestamp spacing. Runs implicitly on every created tub.

- Added TestDataAuthenticity class (8 tests) as confidence gate:
  verifies straight segments have zero yaw, turns have correlated
  gyro/accel, braking produces negative accel_x, all values within
  hardware range, smooth driving has lower noise than jerky.

- Updated all pre-defined profiles and test setups to use
  trajectory-based parameters within realistic RC car ranges.

Total tests: 38 (was 30, added 8 authenticity tests).

https://claude.ai/code/session_01Q6cz4PnXVXz8sPX187A2xM

Author: claude