Expose command dispatch timestamp

[stevenaya]

Summary

• expose the wall-clock time captured immediately before dispatching the latest accepted position target;
• publish the cached command and timestamp only after arm and gripper dispatch calls succeed;
• keep an owned copy of the accepted target so caller-side mutations cannot change driver state;
• invalidate dispatch metadata when a new start session begins.

Why

Downstream nodes need a wall-clock dispatch marker to correlate the command accepted by the driver with external observations. This is a software dispatch timestamp, not motor acknowledgement or physical motion time. The existing monotonic timestamp remains dedicated to velocity limiting.

The boolean send result now belongs to draft PR #23, where safety rejection introduces real False paths.

Validation

• pytest (41 passed)
• ruff check .
• ruff format --check .

[kou]

#28 has been merged. Could you rebase on main?

[stevenaya]

Already rebased on main ^^.

[stevenaya]

Implementation note for the less obvious choices:

• last_command_time_s intentionally uses a monotonic clock for velocity limiting, while last_command_dispatch_timestamp_ns uses wall time for correlation with external observations.
• The wall timestamp is sampled immediately before dispatch, but the public command/timestamp state is committed only after both arm and gripper dispatch calls succeed.
• start() invalidates dispatch metadata from the previous enable session; PR Handle joint delta safety stops without exiting #23 additionally synchronizes the command baseline to measured state.
• np.array(...) keeps an owned snapshot so later caller-side mutations cannot alter last_command.

The boolean return was moved to #23, where safety rejection provides meaningful False paths.