[Bug]: trajectory_metrics silently omits final_pose_speed on low-frequency streams and skews final_pose_unsettled_ratio with unweighted np.mean #700

Version or commit

Current main (3dbf43b)

Problem & Invariant Violation

The documented contract in src/hflow/checks.py:1344-1349 specifies:

"{topic}/final_pose_unsettled_ratio is emitted only when {topic}/mean_velocity is above zero. For a fully motionless episode the ratio is undefined (zero denominator), so only {topic}/final_pose_speed appears. Curation queries joining on final_pose_unsettled_ratio will silently exclude motionless episodes; use final_pose_speed when you need all episodes regardless of mean motion."

This promises that final_pose_speed is always emitted for all episodes. Furthermore, HFlow's motion metrics strictly weight steps by duration (mean_velocity, idle_fraction, _duration_weighted_median) so irregular sampling does not skew results.

However, in checks.py:1424-1433:

1. final_window_mask evaluates (profile.stamps_ns[-1] - profile.stamps_ns[:-1]) / 1e9 <= final_pose_window_s. Because profile.stamps_ns[:-1] is the START of each step, any stream whose sample period exceeds final_pose_window_s (e.g. 1 Hz or 2 Hz streams with the default 0.5s window) has zero steps satisfying the mask. Both final_pose_speed and final_pose_unsettled_ratio are silently omitted despite the robot moving actively. Curation queries joining on final_pose_speed drop these episodes completely.
2. When steps do match, final_speed is calculated using unweighted np.mean(profile.speeds[final_window_mask]). Irregular sampling in the final window distorts final_speed by orders of magnitude and produces severe false positives in final_pose_unsettled_ratio.

Minimal Reproduction

import asyncio
from pathlib import Path
import tempfile
import hflow
from hflow.testing import synthesize_episode, SyntheticEpisodeSpec

with tempfile.TemporaryDirectory() as tmp_dir:
    mcap_path = Path(tmp_dir) / "test.mcap"
    synthesize_episode(
        mcap_path,
        SyntheticEpisodeSpec(
            duration_s=5.0,
            cameras=(),
            joint_hz=1.0,
            joint_jump_at_s=None,
        ),
    )
    with hflow.Episode(mcap_path) as ep:
        result = asyncio.run(hflow.checks.trajectory_metrics(ep))
        assert "/joint_states/final_pose_speed" in result.measurements, (
            "CRITICAL BUG: final_pose_speed was silently omitted!"
        )