fix(dataset): name the malformed YOLO line instead of failing on its shape

[aminmiral]

Description

Fixes #2664. Stacked on #2663, which touches the same function; the diff below is only this change.

yolo_annotations_to_detections appended a class id for every line but only appended a box when the line matched a known shape, with no else. A line of one to four tokens therefore left class_id_list one entry longer than relative_xyxy_list, and the mismatch surfaced far from its cause:

input	before
0 0.5 0.5 0.2	operands could not be broadcast together with shapes (0,) (4,)
a blank line	IndexError: list index out of range
one valid line plus one truncated	class_id must be a 1D np.ndarray with shape (1,), but got shape (2,)
is_obb=True, 6 / 8 / 10 tokens	cannot reshape array of size 5 into shape (2)

The third is what a user sees when one row of an otherwise good file is damaged, and it blames class_id rather than the short line. None of the four name the offending line, so finding it in a large dataset means bisecting by hand.

After:

Invalid YOLO annotation line '0 0.5 0.5 0.2'; expected at least 5 tokens (class id and four box values), got 4.
Invalid YOLO OBB annotation line '0 0.5 0.5 0.5 0.5 0.5'; expected 9 tokens (class id and four corner pairs), got 6.

This matches how the parser already rejects a class id that is not a whole number and a non-numeric trailing token.

Scope

• Five or more tokens are unchanged: boxes, boxes with a trailing confidence or tracker id, and polygons all load as before.
• is_obb=True is documented as nine-token four-corner lines only, so any other count is rejected rather than reaching the reshape.

Type of change

• Bug fix (non-breaking change which fixes an issue)

A file that previously raised one of the errors above now raises a clearer one; nothing that previously loaded stops loading.

How has this change been tested?

Nine new tests in tests/dataset/formats/test_yolo.py: class-only, one, two and three coordinate lines; a blank line; a truncated line among valid ones; and OBB lines that are too few, odd or too many.

Following AGENTS.md §8, with a baseline captured before any change:

baseline   8 failed, 4384 passed, 24 skipped
after      8 failed, 4408 passed, 24 skipped

The failing set is identical in both runs — all eight are pre-existing ImportError: metrics extra is required in this environment. uv run pre-commit run passes on all changed files, mypy included.

Docs

• Docs updated? docs/changelog.md entry added, linking this PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92%. Comparing base (527179a) to head (44a2bad).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #2665   +/-   ##
=======================================
  Coverage       92%     92%           
=======================================
  Files           78      78           
  Lines        11900   11912   +12     
=======================================
+ Hits         11004   11018   +14     
+ Misses         896     894    -2     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Copilot review overview

🟡 Changes recommended

Regression tests do not yet verify the central guarantee that errors identify the malformed annotation row.

Review effort: Balanced
Findings: 1

Open (1)

• Strengthen malformed-line diagnostic assertions · New

What changed in this PR

Improves YOLO dataset-loading diagnostics, alongside the stacked box-extent fixes.

Changes:

• Reject malformed token counts with errors identifying the annotation line.
• Reject negative box extents and normalize exported box corners.
• Add regression tests and changelog entries.

Quality: Code 4/5 · Testing 3/5 · Documentation 4/5
Risk: 2/5 — localized parser and export changes.

File	Description
tests/​dataset/​formats/​test_yolo.py	Tests malformed lines, box extents, and export ordering.
src/​supervision/​dataset/​formats/​yolo.py	Validates annotation structure and normalizes box extents.
docs/​changelog.md	Documents diagnostic and box-handling changes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[aminmiral]

Addressed the Copilot note on the malformed-line tests.

They matched only token, which was too loose to protect the thing this PR is for: I stripped the quoted row and the token counts out of both error messages and all nine tests still passed. Each test now matches the full message, so the offending row and the expected/actual counts are both pinned, and the mixed valid/truncated case asserts the truncated row is the one named. With the diagnostic removed all nine now fail.

Also merged develop in. All checks green.