[shashwat-deep]

Follows Pratyush's review of #10. Stacked on #17 (eval-hashes); once #17 is merged, this PR moves to main.

What changes

• eval/build_manifest.py leaves out any image whose file carries an EXIF orientation other than 1, or EXIF data that Pillow cannot read, and takes the next image in its split's ranking, checking that one in turn. The browser turns an image by its EXIF orientation, and Open Images' boxes and Rotation value do not say whether they allow for it. No image of the current draw is affected: eval/manifest/ is unchanged.
• It also prints how many drawn files are larger than 1,024 px on the long side, carry EXIF data, or carry an ICC profile. For the current draw: 2, 2 and 430, and 0 left out. The two large files are the tune images 6ae23606d22bad2d (3780 x 6111) and cce0d9c5636f322e (6000 x 4000), both with EXIF, IPTC, XMP and ICC data, and Rotation 0.
• run_detect.py, render_copies.py and check_colour.py refuse such a file before the browser starts, whatever the transform, and transforms.upright() refuses it too.
• The runner gives an image stored turned its own orientation tag in place of any EXIF data the file has, where it used to refuse a turned file with EXIF data. So every image the builder keeps can be served. The 15 turned images of the current draw have no EXIF data and are served byte for byte as before.
• The orientation is read as Chromium reads it (transforms.exif_orientation()): from the first EXIF segment only, and not from XMP's tiff:Orientation, which Pillow's getexif() also reports. A test checks this in Chromium 156.
• docs/eval-protocol.md (part B: Data, Exclusions, Orientation and the tiled setting) and eval/README.md no longer say that every CVDF copy is at most 1,024 px with all metadata removed, and list the new exclusion. No evaluation run has been made yet, so this is not a deviation from the protocol.

Checks

• .venv/bin/python -m pytest eval/tests: 104 passed.
• Removing each new rule in turn (the builder's exclusion, the runner's refusal, the as_stored condition, the EXIF-only reading, the replacement of EXIF data, the unreadable case, check_colour.py's check) makes a test fail.
• eval/build_manifest.py on the real metadata leaves eval/manifest/ byte for byte unchanged, and all 1,000 images pass check_orientations().
• /code-review (max): 9 findings, all fixed in the second commit.

🤖 Generated with Claude Code

[06pratyush]

Approving. I checked:

• the exclude-and-replace loop in build_manifest.py: each image left out for its EXIF orientation is replaced by the next one in its split's ranking;
• the refusals in run_detect.py, render_copies.py, check_colour.py and transforms.upright();
• the runner replacing the EXIF data of an image stored turned with its own orientation tag;
• the test that Chromium turns an image by its first EXIF segment only.

The eval tests pass on Windows at the top of the stack (104, none skipped).