[Arthur031221]

Learners preparing their own pet images through COURSE_PET_IMAGES can have a PNG with valid pixels but malformed EXIF accepted into the split, then see decoding fail in pet_dataset.

prepare_pets checked image structure and decoded RGB pixels without applying the loader's EXIF orientation transform, so malformed metadata escaped validation. Preparation now applies ImageOps.exif_transpose before RGB conversion and records malformed EXIF through the existing rejection path. The regression uses a synthetic PNG with valid pixels and malformed EXIF, checks rejection, and decodes all three splits. The Microsoft dataset was not inspected for malformed EXIF.

On Linux/Python 3.12.3, .venv/bin/python -m unittest discover -s tests -q passed 18 tests. Basic and deep checks passed 15 and 4 scripts respectively, with matching notebook counts. Sync and docs checks passed. Python 3.11.15 also passed the basic checks, with 9 optional deep tests skipped. The regression fails on d430ac5, passes with the fix, and fails again when only the source change is reverted.