Reviewed the current implementation against main (7bcd2144..bb06b5ed). The change is narrowly scoped and the targeted atomic-action tests pass, but I think this needs fixes before merge.

Blocking / should fix

1. Multi-waypoint trajectories do not guarantee exact visits to intermediate waypoints

   • embodichain/lab/sim/atomic_actions/trajectory.py:336
   • embodichain/lab/sim/atomic_actions/trajectory.py:360

   The implementation passes all keyframes to interpolate_with_distance() and resamples the entire path to a fixed count. That preserves path order, but it does not guarantee intermediate waypoints appear as emitted trajectory samples. For example, keyframes [0, 1, 3] with 5 samples become [0, 0.75, 1.5, 2.25, 3], so waypoint 1 is skipped. This conflicts with the new “visits every waypoint” API contract and can miss clearance/safety waypoints.

   Suggested fix: interpolate segment-by-segment while preserving each keyframe boundary, or explicitly insert exact keyframes into the sampled output. Also validate that sample_interval can accommodate all required keyframes.

2. Empty multi-waypoint targets silently succeed as no-ops

   • embodichain/lab/sim/atomic_actions/trajectory.py:88
   • embodichain/lab/sim/atomic_actions/trajectory.py:162
   • embodichain/lab/sim/atomic_actions/actions.py:239

   (n_envs, 0, 4, 4) and (n_envs, 0, dof) pass validation. Planning then prepends only start_qpos, returning a successful stationary trajectory. These should be rejected with a clear error, and tests should cover empty waypoint tensors.

3. EndEffectorPoseTarget is now ambiguous for Place

   • embodichain/lab/sim/atomic_actions/core.py:73
   • embodichain/lab/sim/atomic_actions/actions.py:687

   EndEffectorPoseTarget now documents the 4D multi-waypoint form, and Place accepts the same target type, but Place.execute() still expects a single pose and will fail later through a generic shape check. Either reject 4D targets in Place with a clear error, split the target type, or implement multi-waypoint place behavior.

4. Tests mock out the behavior that most needs protection

   • tests/sim/atomic_actions/test_actions.py:174
   • tests/sim/atomic_actions/test_actions.py:287
   • tests/sim/atomic_actions/test_trajectory.py:252

   The new tests assert that keyframes are passed into the interpolator, but not that the returned trajectory actually preserves intermediate waypoints. Add a real interpolation regression test with a simple uneven path and assert intermediate waypoints are emitted.

Minor

• Public docs still list only single-target shapes:
  • docs/source/overview/sim/atomic_actions/index.md:97
  • docs/source/overview/sim/atomic_actions/builtin_actions.md:29

Verification

Ran:

pytest -q tests/sim/atomic_actions/test_trajectory.py tests/sim/atomic_actions/test_actions.py

Result: 45 passed in 36.31s.