NMS / NMM and OBB metrics call oriented_box_iou_batch with the raw box
coordinates, so the rasterization canvas grew with the source image
resolution — HD/4K inputs allocated (N, ~4000, ~4000) uint8 masks, easily
gigabytes for a few hundred detections.

IoU is invariant under translation and uniform scaling, so we shift boxes
to the origin (canvas no longer pays for dead space when boxes cluster in
a corner) and shrink them when the bounding region exceeds a 1024-pixel
cap. Memory is now O(N * 1024^2) regardless of input resolution.

Tests cover both transforms via a single parametrized case.

- [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>

- [resolve roboflow#5] oriented_box_non_max_suppression + oriented_box_non_max_merge:
  add Examples: doctest blocks (project convention for all public APIs).
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- [resolve roboflow#7] oriented_box_iou_batch: update summary/Returns from "Pairwise IoU"
  to overlap_metric-aware wording, matching box_iou_batch convention.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- [resolve roboflow#8] oriented_box_iou_batch + public entry-points: add ndim/shape
  validation before reshape to catch flat (N,8) and wrong-ndim inputs early.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- [resolve roboflow#9] oriented_box_non_max_merge: add assert 0 <= iou_threshold <= 1,
  matching the guard present on oriented_box_non_max_suppression and siblings.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested

[resolve group] PR roboflow#2303 — items 5, 7, 8, 9

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

- [resolve roboflow#4] test_with_nmm_falls_back_to_box_nmm_without_obb_data: regression
  guard that non-OBB Detections still collapse via box NMM path.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- [resolve roboflow#10] boundary tests: empty/single, iou_threshold=0.0/1.0, IOS metric
  end-to-end NMS, and OBB NMM empty/single — converted threshold test to
  parametrize(expected_keep) removing if-branch from test body.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested
- Added test_with_nmm_obb_merged_xyxy_matches_winner_aabb verifying the
  item roboflow#3 contract: merged xyxy == winner's AABB, not union of all members.

[resolve group] PR roboflow#2303 — items 4, 10

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

…low#2303

- [resolve roboflow#12] add UnReleased entry: with_nms + with_nmm now use OBB IoU when
  data["xyxyxyxy"] present; callers on old AABB path should strip that key;
  threshold values may need recalibration.
  Challenge: evidence=VALID suggestion=VALID resolution=as-suggested

[resolve group] PR roboflow#2303 — item 12

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

- oriented_box_non_max_suppression: fenced block → >>> doctest, asserts keep == array([ True, False])
- oriented_box_non_max_merge: fenced block → >>> doctest, asserts len(groups) == 1

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

…rize

test_drops_true_duplicates + test_is_class_aware → test_suppression_is_class_aware
parametrized on (class_id_b, expected_keep): same class suppresses, diff class keeps both

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

- test_respects_overlap_metric → test_overlap_metric_determines_suppression parametrized on (overlap_metric, expected_keep); removes dual-act
- TestDetectionsWithNms + TestDetectionsWithNmm shared tests → TestDetectionsObbDispatch parametrized on method name; TestDetectionsWithNmm retains only xyxy-specific test

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