v0.2.0

• Final whole-arm zero convergence corrects drift after intermediate zero moves;
  all eight joints must settle together before disabled-pose checks and writes.

• Calibration moves now use slow, bounded target correction for static position
  error; actual-zero tolerance and manual-hold gains are unchanged. Failure reports
  include target/goal/feedback history captured before disable.

• Calibration move/seek speeds are now 20/10 degrees per second, with ramped
  upstream per-action Kp/Kd settings. Inactive-joint failures include the offending
  feedback.

• Single-arm v2 automatic calibration over Classic CAN, available through the
  Python API, CLI and a simplified arm/action selector in the follower panel.

• Panel assumes MIT when RID10 times out, without an advanced checkbox; explicit
  non-MIT readback is still rejected. CLI/API retain the explicit timeout option.

• Bounded stop seeking, computed-zero positioning, cancellation/disable cleanup,
  and disabled-pose checks with per-motor hardware zero write reports.

• Offline virtual-stop and failure-path tests. Both arms completed zero writes
  (right 139 s, left 126 s); all 16 angle readbacks were unchanged after one user
  power cycle. Absolute accuracy and repeatability remain unquantified.

• Existing hold/leader APIs remain compatible; native SDK and CAN bitrate are
  unchanged. See docs/calibration.md for operation and recovery.