[gbradham]

Summary

Session Review said "No poses found" for every run recorded with a live pose model. On the Maimu RTPP Day1 recordings it also played the overlay video instead of the raw recording, and named the session …_annotated_<ts>. Three causes, all fixed here:

• The live keypoints were never read. SessionView.from_recording used the tracked centroid as a one-keypoint pose and ignored the <base>_keypoints.csv the tracking logger writes beside the tracking CSV. It now loads that file:
  • It's in long format (frame, object_id, keypoint, x, y, confidence), and only the subject's (frame, object_id) rows are kept, so a second animal never mixes in.
  • Blank x/y, which the logger writes for an undetected keypoint, becomes a gap.
  • The logger's keypoint labels are kept in order: the CV-settings bodypart names, or indices when none were set.
  • It falls back to the centroid when the file is missing or unusable.
• The overlay video was taken for the raw recording. discover() and session_id_for() only recognised <stem>_annotated.mp4, but VideoRecorder writes <name>_annotated_<ts>.mp4. The overlay copy was therefore classified as the raw video and won as the newest file.
• Opening a video used the wrong CSV as poses. find_pose_csv() treated <video stem>.csv as a hand-placed pose file, but for a live recording that's the DataRecorder's sensor CSV. Files that start with a # GLIDER header are now skipped.

New helpers in glider.analysis._io: is_annotated_video, raw_video_stem, keypoints_beside. Session exposes tracking_path. The Session tab's wording now says whether the poses are the live keypoints or the centroid.

Not in this PR

A folder holding several runs still opens as one session, the newest. GLIDER saves runs flat, so a folder like Day1 (9 runs) shows one. Each recorder also timestamps its own filename, so one run's files can differ by a second (…172144.mp4 vs …172145_keypoints.csv). Both are separate follow-ups.

Test plan

• On the real Day1 share, read-only: the newest run opens with 7 keypoints on all 14,079 frames, the raw video, and session id Maimu_RTPP_1_20261020_172144. find_pose_csv on an older run's video no longer returns its data CSV.
• New unit tests:
  • keypoints become the pose
  • bodypart names kept in logged order
  • an undetected keypoint becomes a gap
  • only the subject's keypoints kept
  • unreadable or empty keypoints fall back to the centroid
  • the recorder's annotated naming is recognised (discover, raw_video_stem)
  • keypoints_beside
  • a GLIDER CSV is never taken for poses, and doesn't shadow a batch output
• Full pytest tests (offscreen, PYTHONPATH=src): 6,820 passed, 20 skipped
• ruff check . and black --check . clean