v0.2.0 — 可组合观测段与传感器列表

Breaking

• The default observation is no longer 494 values. ObsSpec.default(action_dim, path_points=30)
  builds odom + path + goal + act, which is 128 values on the Ackermann and
  differential-drive chassis and 129 on the mecanum (the extra value is the
  third action component). The old 494-value layout is now an explicit choice:
  ObsSpec assembled from ObsLayout(2, 30).as_spec(), passed as
  ChassisSimEnv(..., obs_spec=...).
• SensorConfig is gone; sensors are a list of specs. Build
  Lidar / Imu / WheelOdom instances and pass them as
  ChassisSimEnv(..., sensors=[...], sensor_seed=0) or
  SimpleBridge(chassis, sensors=[...], sensor_seed=0). sensors= and an explicit
  bridge= are mutually exclusive.
• chassis_sim/bridges/sensors.py was deleted. The old single-config sensor
  bridge no longer exists; sensor state lives in chassis_sim.sensors and in the
  bridge's own get_sensor_reading().
• Sensor readings stay in info (info["scan"], info["imu"], info["wheel"])
  and never enter the observation vector unless a LidarSegment observation segment
  explicitly names the sensor. Reset now carries the same readings as step()
  (the bridge produces a real first frame inside reset()), and still returns
  {} when no sensor is configured.
• chassis_sim.obs.Lidar → LidarSegment（与 chassis_sim.sensors.Lidar 重名）.
  The observation segment is now chassis_sim.obs.LidarSegment; the sensor spec
  keeps the name chassis_sim.sensors.Lidar and there is no compatibility alias.

Added

• chassis_sim.obs: composable ObsSpec with segments Odom, Path,
  Goal, Act, Err, GoalSig, PathValid, PathProgress, LidarSegment, and
  Custom; plus ObsSpec.default(), ObsSpec.resolved(),
  ResolvedObsSpec.total_dim, ResolvedObsSpec.describe(), and
  env.obs_spec_describe() returning name -> slice.
• chassis_sim.sensors: list-style specs Lidar(name, mount, fov, n_rays, max_range, range_min, noise_std, quant_step, latency_steps, dropout_prob),
  Imu(name, noise_std), WheelOdom(name, track_width, noise_std). Each sensor
  owns a random stream derived from (sensor_seed, name), so adding or removing a
  sensor never shifts the noise of the others.
• Missing-frame contract: the bridge produces a real first LiDAR frame right after
  reset(); a delayed or dropped frame makes the Lidar observation segment repeat
  the last valid scan (zero-order hold), and max_range is used only before the
  first valid frame. env.reset() clears the hold cache. scan_meta["valid"] /
  scan_meta["dropped"] remain the only validity signal.
• examples/gallery.py, which regenerates the five clips and six scenario stills
  used by the README gallery, and the README gallery section itself.
• README rewritten for 0.2.0: the observation contract, the explicit 494 preset,
  the sensor spec list, the zero-order hold contract, a regenerated evaluation
  matrix, and executable, output-checked code blocks throughout.

Fixed

• The README documented PhysicalCommandBridge.step() as accepting physical
  commands. It accepts the normalized action and calls
  inner.step(chassis.physical_action(action)); to_physical() is the pure
  mapping. Documentation corrected, behavior unchanged.
• Documentation no longer refers to SensorConfig, LidarParams,
  bridges/sensors.py, the side= argument, or get_lidar_front, and no longer
  describes a fixed 494-value observation outside the explicit preset section.