[resolve #6] /review finding by doc-scribe: add note to force_masks docstring
that it has no effect when is_obb=True, preventing caller confusion.

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Add filtered_empty_with_mask parametrize case (same bug applies to mask field)
- Add explicit ids= for readable pytest output
- Add one-line docstring to test function

[resolve roboflow#5] /review finding by qa-specialist (report: .temp/output-review-fix-is-empty-tracker-id-2026-04-13.md)
[resolve roboflow#6] /review finding by qa-specialist (report: .temp/output-review-fix-is-empty-tracker-id-2026-04-13.md)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

[resolve #6] /review finding by sw-engineer (report: .temp/output-review-from-inference-rle-mask-2026-04-14.md):
"Broad exception swallowing at internal.py:107 silently drops RLE decode failures with no warning..."

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Parametrized crash-regression test for 1/2/3/4 unique anchor
  positions (covers the splprep threshold boundary at len==4).
- Fallback-matches-raw assertion for <4 unique positions (verifies the
  new else branch uses unique_xy, same as smooth=False raw polyline).
- Single-frame smooth=True no-crash test (empty np.diff edge case).
- smooth=False stationary regression guard (ensures refactor left the
  non-smooth path untouched).

[resolve roboflow#5] /review finding by qa-specialist (report: boundary conditions untested)
[resolve roboflow#6] /review finding by qa-specialist (report: single-frame smooth=True untested)
[resolve roboflow#8] /review finding by qa-specialist (report: smooth=False regression guard)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Attacker-controlled file_name from COCO JSON can contain embedded null
bytes or other invalid chars that cause Path().resolve() to raise OSError
or ValueError. Catch and re-raise as a descriptive ValueError that names
the offending file_name rather than leaking a low-level OS error.

[resolve roboflow#6] Review comment by @Copilot (gh):
"Path(...).resolve() can raise for embedded nulls — re-raise as descriptive ValueError"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- Replace old C-order example [1, 3, 2] with F-order [1, 2, 1, 1, 1] for
  the same 2×3 mask, matching _rle_encode and _rle_decode examples
- _rle_area is order-agnostic but consistent examples across all three
  helpers reduce reader confusion

[resolve #6] /review finding by sw-engineer (report: .temp/output-review-rle-coco-2026-04-22.md):
"_rle_area docstring example still uses old C-order RLE for consistency"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Replace O(height) Python for-loop with np.diff + np.flatnonzero to find
run boundaries, eliminating per-row Python overhead in the RLE scale path

[resolve #6] /review finding by sw-engineer (report: .temp/output-review-rle-resize-2026-04-22.md):
"_rle_scale_col re-encodes output column with a Python for loop instead of vectorized np.diff/np.flatnonzero"

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…ess output and log stderr on failure

- Add `-loglevel error -nostats` so ffmpeg only writes actual errors to stderr
  (eliminates progress/stats spam that would buffer in PIPE indefinitely)
- Decode `result.stderr` and include it in the warning when ffmpeg exits
  non-zero, so failure messages surface diagnostically instead of being discarded

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…ng and edge cases

[resolve #6] /review finding by foundry:qa-specialist (report: .reports/review/2026-05-22T05-56-29Z/review-report.md)

---
Co-authored-by: Claude Code <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <codex@openai.com>

…etections

Three additions:
- Assert cv2.imwrite return value in _tiny_detection_dataset helper so
  codec failures produce explicit errors rather than cryptic downstream ones
- Fix _tiny_detection_dataset xyxy shape for dets_per_image=0 (was (0,),
  now (0,4) via reshape — required by Detections validator)
- test_save_coco_annotations_annotation_image_id_references_correct_image:
  verifies every annotation.image_id references a known image id after
  chaining with a non-default starting_image_id
- test_save_coco_annotations_zero_annotation_images: pins behaviour for
  images with zero detections — image ids advance, annotation list empty

[resolve roboflow#6] /review finding by foundry:qa-specialist (report: .reports/review/2026-05-22T18-44-11Z/review-report.md):
[resolve roboflow#7] /review finding by foundry:qa-specialist (report: .reports/review/2026-05-22T18-44-11Z/review-report.md):
[resolve roboflow#10] @Copilot (gh comment #3290385653):

---
Co-authored-by: Claude Code <noreply@anthropic.com>

…_inference

No test exercised the duck-typed .dict() object path (the documented Inference SDK
usage) for empty predictions. Add test with minimal fake SDK object to confirm the
dtype fix applies through the .dict() conversion branch.

[resolve roboflow#6] /review finding by foundry:qa-specialist (report: .reports/review/2026-05-27T06-06-27Z/review-report.md)

---
Co-authored-by: Claude Code <noreply@anthropic.com>

- [resolve roboflow#3] with_nmm: for OBB merge groups with >1 member, set merged
  xyxy from winner's OBB corners instead of union AABB of all members.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- [resolve roboflow#6] with_nms + with_nmm docstrings: document three-path dispatch
  order (mask → OBB → AABB) so callers know OBB is now routed differently.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- [resolve roboflow#11] with_nmm: cast OBB to float32 (was float64), aligning with
  with_nms; relax _group_overlapping_oriented_boxes annotations to np.floating.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested

[resolve group] PR roboflow#2303 — items 3, 6, 11

---
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>