fix(checks): duration-weight final_pose_speed and retain on sparse streams

[shobhitagnihotri69]

Closes #700

Summary

• Retain final_pose_speed on low-frequency / sparse streams: final_window_mask previously matched against step start timestamps (profile.stamps_ns[:-1]), silently dropping final_pose_speed and final_pose_unsettled_ratio whenever the step period exceeded final_pose_window_s (e.g. 1 Hz streams with the default 0.5s window). Steps ending after the window start now represent the settling motion, with a fallback to the last measured step.
• Duration-weight final_pose_speed: previously computed using unweighted np.mean(profile.speeds[final_window_mask]), which violated HFlow's sampling-invariance principle and skewed under irregular sampling or jitter. final_speed is now weighted by profile.step_durations_s, matching mean_velocity.
• Identity on whole-window episodes: when the final window covers the full episode, duration-weighting guarantees final_pose_speed matches mean_velocity identically and final_pose_unsettled_ratio evaluates to 1.0.
• Regression coverage: added tests to tests/test_checks.py verifying that low-frequency streams emit final_pose_speed and that duration weighting matches mean_velocity on full-window episodes.

Testing

• uv run ruff check src/hflow/checks.py tests/test_checks.py (passes)
• uv run ruff format --check (passes)
• uv run ty check (passes)
• uv run pytest tests/test_checks.py -k trajectory_metrics (7 passed)
• uv run pytest -m "not slow" -q (2483 passed, 49 skipped)

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Changes how trajectory metrics weight motion data.

This PR appears safe to merge; no blocking issues remain.

What we checked:

• Sparse streams retain final speed: A finite final step overlaps a positive final window. Its speed is still included even when samples arrive less often than the window length.

Summary

This PR weights final_pose_speed by each step’s time inside the final window. Sparse streams retain the metric when their final motion can be measured.

• Final-pose speed reflects time spent moving near the episode's end.

_{Reviews (4) · Last reviewed commit: "fix(checks): omit final_pose_speed when ..." · Reviewed by Greptile}

[kstonekuan]

Thanks, the overlap weighting is right and fixes both halves of #700. One change: drop the fallback to the last measured step. With overlap weighting, a sparse stream's final step already overlaps the window, so the fallback only fires when the final steps are unmeasurable (non-finite). Then it reports speed from earlier in the episode as the final pose. Omitting the measurement is the honest answer there. Please also add a test for that case.

[kstonekuan]

LGTM, merging.

[shobhitagnihotri69]

Thanks very much for merging this pr looking forward to do more contribution