Report file, line and reason for skipped pose label lines (#873)

[EHxuban11]

Closes the logging request left open by #873. Parsing rules are unchanged: the
same lines are still rejected.

What changed

• YOLOPoseDataset names the label file, the line number and the actual reason
  for each skipped line, quoting the first 5 and counting the rest.
• A non-numeric class id or coordinate is now reported as such. Previously one
  except ValueError covered both the length check and number parsing, so every
  rejection printed "field count does not match", including ones that were not
  about field count.
• A wrong-length line says how many keypoints it carries and whether it matches
  the other kpt_shape keypoint dim.
• PoseValidator._build_coco_gt_from_yolo dropped malformed ground-truth lines
  with no message at all, so reported mAP was computed over fewer objects than
  the user expected with nothing in the log. It now warns. Not part of Pose dataset loader drops label lines without saying which ones #873.
• docs/dataset_schema.md and the module docstring state the accepted xy /xyv
  layouts and that a bad row is skipped alone, not the whole file.

Review notes

• bbox and keypoints are now slices of one array instead of two separate
  np.array calls; they are disjoint views, and both call sites copy.
• MAX_REPORTED_BAD_LINES = 5 caps log volume on a dataset with many bad rows.
• format_bad_label_lines is imported by the validator, hence public.

Verification

• New tests/unit/test_pose_label_diagnostics.py (9 tests): valid [K, 2] and
  [K, 3] layouts keep every line, the warning names file and line, a numeric
  error is not reported as field count, quoting is capped.
• New test in tests/unit/test_pose_validator.py for the dropped ground truth.
• pytest tests/unit -k "pose or dataset or validat": 876 passed, 29 skipped.
• Not verified: no real training or validation run on a GPU.

Code provenance

Original code written for this PR. No third-party code was ported, adapted, or
derived. The changes are confined to error reporting in
libreyolo/data/pose_dataset.py and libreyolo/validation/pose_validator.py,
their tests, and documentation.

Confidence Score: 5/5

The PR appears safe to merge; no actionable failures remain in the reviewed changes.

Summary

This PR adds bounded, line-specific diagnostics when malformed YOLO pose labels are skipped during dataset loading or validation.

• Reports the label file, line number, and actual parsing failure.
• Limits quoted examples while retaining the total rejected-line count.
• Warns when malformed validation ground truth is excluded from pose metrics.
• Documents the accepted xy and xyv layouts and row-level rejection behavior.
• The previous non-finite class-ID failure is fixed by translating OverflowError into the parser's existing ValueError contract.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Pose label row] --> B{Field count valid?}
    B -->|No| E[Record file, line, and reason]
    B -->|Yes| C{Class and coordinates parse?}
    C -->|No| E
    C -->|Yes| D[Add pose annotation]
    E --> F[Skip only malformed row]
    F --> G[Emit bounded warning summary]
    D --> H[Training or pose evaluation]

Loading

_{Reviews (2) · Last reviewed commit: "Skip non-finite pose class ids instead o..."}