[BSAIPF]

Image.new_from_array() drops interpretation when it falls back to an object's __array__ method. For example, a float32 PyTorch RGB tensor with interpretation='auto' produces srgb, while passing its NumPy array directly produces scrgb. An explicit interpretation such as lab is also ignored.

Forward interpretation through the fallback call. Add 15 regression/control cases covering array-only objects, RGB and grayscale data, uint8/uint16/float32, and real PyTorch tensors. The tests also check unchanged pixel values and the default interpretation.

Validation on macOS ARM64, Python 3.12.14, libvips 8.18.7 (API mode):

• New cases against the original implementation: 9 failed, 6 passed.
• Full pytest -q -ra: 78 passed, no skips (NumPy, Pillow, SciPy and PyTorch installed).
• flake8 . passed.
• Both CI mypy commands passed: mypy pyvips/*.pyi --no-error-summary and mypy examples/*.py --no-error-summary.
• git diff --check passed.

Prepared with Codex assistance.

[kleisauke]

Please do not use LLMs to generate code or PR descriptions when contributing to libvips:
https://github.com/libvips/libvips/blob/master/CONTRIBUTING.md#ai-policy

Evidence

The PR description mentions the use of OpenAI Codex, as confirmed by the audit log:

user_agent	GitHub CLI 2.102.0 Agent/codex

See also: https://agentscan.tools/user/BSAIPF.

[jcupitt]

Hi @BSAIPF,

I made a tiny, non-LLM PR for this fix and credited you:

#1430