fix(detection): keep from_inference aligned on partial masks

[kounelisagis]

Bug

process_roboflow_result appends a mask only for predictions that carry one (RLE or polygon), but appends xyxy, confidence and class_id for every prediction. A result that mixes masked and box-only predictions therefore ends up with a mask array shorter than the boxes, and Detections.from_inference raises:

ValueError: mask must be a 3D np.ndarray with shape (2, H, W), but got shape (1, H, W)

This shows up with segmentation results where one polygon is empty/degenerate and falls back to a box, so part of the batch has masks and part does not.

This is the same shape of issue as #2353 (partial tracker_id), in the same function; the mask path was the remaining case.

Fix

Append None for box-only predictions to keep the list aligned, and build the mask array only when every prediction has a mask, otherwise drop masks (with a warning) to preserve alignment with xyxy. Fully-masked and mask-free results are unchanged. Missing entries are detected by identity (mask is None) since None in masks would compare numpy arrays element-wise and raise.

Tests

Added an end-to-end from_inference test for a mixed masked/box-only batch, and updated the existing process_roboflow_result case that previously documented the misalignment as a known limitation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83%. Comparing base (15f56de) to head (70ded47).
⚠️ Report is 1 commits behind head on develop.

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #2362   +/-   ##
=======================================
  Coverage       83%     83%           
=======================================
  Files           69      69           
  Lines         9616    9622    +6     
=======================================
+ Hits          7941    7947    +6     
  Misses        1675    1675           

🚀 New features to boost your workflow:

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

[Copilot]

Pull request overview

This pull request fixes a robustness issue in the Roboflow/Inference connector path by ensuring mask handling stays aligned with xyxy when a batch mixes masked and box-only predictions. It updates process_roboflow_result to avoid producing misaligned mask arrays that previously caused Detections.from_inference to raise on mixed batches.

Changes:

• Ensure process_roboflow_result appends None for box-only predictions and drops masks (with a warning) when only a subset of predictions includes mask data.
• Update the existing unit test that previously documented the misalignment as a known limitation to reflect the new aligned behavior.
• Add an end-to-end Detections.from_inference regression test covering a mixed masked/box-only batch.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/supervision/detection/utils/internal.py	Keeps mask collection aligned by inserting None for box-only detections and dropping masks on partial-mask batches to prevent shape errors.
tests/detection/utils/test_internal.py	Updates the expected output for a mixed RLE + box-only batch to confirm masks are dropped (not misaligned).
tests/detection/test_core.py	Adds an end-to-end regression test asserting mixed mask presence does not crash and results in mask is None.