Index and display .tif, .avif and .jfif images

[lstein]

Fixes #364.

Problem

IMAGE_EXTENSIONS contained .tiff but not .tif (the spelling scanners, Photoshop and Windows tools produce), .avif or .jfif. An album of .tif scans indexed as silently empty, and the serving guard 403'd those suffixes.

Changes

• Taxonomy: add .tif, .avif and .jfif to IMAGE_EXTENSIONS, and update the pinned literal in test_media_types.py.
• Pillow floor: Pillow>=11.3.0. I checked the wheels: 11.1.0 and 11.2.1 report features.check('avif') == False, and 11.3.0 reports True. The lock is unchanged at 12.2.0.
• HEIF opener: register_heif_opener() moves from an import side effect in embeddings.py into media_types.py. is_image("x.heic") now implies PIL can open it even for callers that import only media_types.
• Displaying TIFF: only Safari renders TIFF, so a .tif that now indexed would still be a broken image in Chrome and Firefox. /images/ now converts TIFF to PNG as it already did for HEIC (BROWSER_CONVERTED_EXTENSIONS). The conversion:
  • runs in asyncio.to_thread, so a large scan doesn't stall other requests;
  • converts CMYK through the embedded ICC profile when there is one, and drops the stale profile otherwise;
  • rescales I;16B, I and F images to 16-bit instead of clipping them to a blank white or black page;
  • uses compress_level=1.
• Downloads: single-item download now fetches ?original=1, so a saved .tif or .heic is the user's file and not the flattened PNG. This also fixes the same existing problem for HEIC.
• Docs: the supported-format lists are updated.

Tests

The new tests/backend/test_image_formats.py covers:

• every image suffix, generated at test time, lower- and upper-case, is picked up when indexing;
• every suffix, plus a CMYK TIFF, is served in a form PIL decodes, and converted formats come back as image/png;
• ?original=1 returns the file's exact bytes;
• deep TIFF modes aren't clipped;
• in a fresh interpreter, every is_image() format opens with embeddings never imported;
• a completeness check fails if a suffix is added without a test image.

Full suite: 1224 backend and 953 frontend tests pass; ruff, eslint and prettier are clean.

Review

A fresh-context adversarial review found the download regression and the deep-TIFF clipping. Both are fixed here. It also flagged:

• Uncached conversion: about 9 s per request for a 24 MP scan, with no ETag. This is deferred to Cache (and downscale) TIFF/HEIC conversions served by /images/ #414.
• .jfif MIME type: Python's standard mimetypes table has no .jfif entry. This is cosmetic, because <img> sniffs the bytes.

🤖 Generated with Claude Code