fix: degrade gracefully when OpenCV is present but unusable

[hiroki-abe-58]

Summary

Related to #1. Thanks to @greatweir-star for the detailed report and reproduction.

detect_faces only caught ImportError. An OpenCV that imports but cannot be used raised into pipeline.act() and ended an otherwise deterministic lattice/granule run with outcome: "failed". Examples are OpenCV 5.x without a top-level CascadeClassifier, a build without cv2.data or cascade files, or a cv2.error at run time.

This PR keeps face protection best-effort and keeps the degraded state diagnosable:

OpenCV state	before	after
not installed	[], cv2_unavailable	unchanged
import fails inside cv2 (e.g. DLL / ABI error)	[], cv2_unavailable	[], cv2_error:ImportError
no CascadeClassifier / no cv2.data	AttributeError, run failed	[], cv2_error:AttributeError
no cascade file can be loaded	[], haar_cascade_no_face	[], cv2_error:FileNotFoundError
error at run time (detectMultiScale, ...)	exception, run failed	[], cv2_error:<type> (cv2_error:error for cv2.error)
working OpenCV	boxes, haar_cascade / haar_cascade_no_face	unchanged

Implementation notes:

• The OpenCV calls move unchanged into _haar_cascade_boxes().
• detect_faces() catches Exception (not BaseException) only around the cv2 import and that helper. Box merging stays outside the try.
• The check is not limited to CascadeClassifier: any failure in the OpenCV part degrades.
• No change to pipeline.py, thresholds or signal-processing behaviour.

Tests

tests/test_face_detection.py covers:

• each broken case, using small fake cv2 modules (via monkeypatch);
• a fake cv2 package whose import raises;
• a working fake, to check that merged boxes are unchanged;
• a lattice pipeline.run with a broken cv2, which must still finish with diagnose → notch → verify → finish and verify passed.

Verification

Local only: Windows 11, Python 3.12.13, numpy 2.5.3.

• New tests on unmodified main (865a148): 6 failed, 2 passed. The 2 that pass are the unchanged paths.
• This branch, without OpenCV: pytest 23 passed.
• This branch, with opencv-python-headless 4.14.0.94 (.[face]): pytest 23 passed.
• ruff check . is clean.
• With real OpenCV 4.14, detect_faces returns identical boxes and status on unmodified main and on this branch for 7 local test images (one with a detected box).
• I did not install a real OpenCV 5.x. The incompatible builds are reproduced with fake modules only.
• Not run locally: macOS, Python 3.11 / 3.13 (left to CI).

🤖 Generated with Claude Code