feat(annotate_subtasks): only downsample frames larger than target_size

[WilliamYue37]

What this does

_resize_and_center_crop in src/opentau/scripts/annotate_subtasks.py previously upsampled small frames (e.g. 224×224) up to target_size × target_size (default 448×448) and center-cropped. Upsampling adds no information — it just inflates image-token cost on the Anthropic / Gemini API and slows the request.

This PR changes the behavior so the function:

• Downsamples + center-crops frames whose shorter side is strictly greater than target_size (unchanged path for typical 30–50 fps recordings shot at >= VGA).
• Returns frames whose shorter side is <= target_size untouched at their native resolution (no upsample, no crop).

Module docstring, function docstrings, and the --target-size CLI help are updated to match. (🗃️ Feature)

How it was tested

• Updated tests/scripts/test_annotate_subtasks.py::TestResizeAndCenterCrop:
  • Renamed test_smaller_than_target_is_upscaled → test_smaller_than_target_passes_through and inverted its assertion (224×224 in → 224×224 out).
  • Added test_non_square_smaller_than_target_passes_through (300×400 in → 300×400 out).
  • Added test_short_side_at_target_passes_through_without_cropping (448×600 in → 448×600 out).
• Existing landscape (640×480) / portrait (480×640) / already-square / arbitrary-target cases still pass — those frames have shorter side > target so the downsample-and-crop path is exercised.
• pre-commit run --files src/opentau/scripts/annotate_subtasks.py tests/scripts/test_annotate_subtasks.py → all hooks pass.
• pytest tests/scripts/test_annotate_subtasks.py -m "not gpu" -x → 48 passed.

How to checkout & try? (for the reviewer)

git checkout feat/annotate-subtasks-no-upsample
pytest -sx tests/scripts/test_annotate_subtasks.py::TestResizeAndCenterCrop

Checklist

• I have added Google-style docstrings to important functions and ensured function parameters are typed.
• My PR includes policy-related changes.
  • If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.

🤖 Generated with Claude Code

[claude]

[claude-review] summary for commit 4948551

No blocking issues found.

Verified:

• Branch logic in _resize_and_center_crop (src/opentau/scripts/annotate_subtasks.py:156-178) matches the docstring: pass-through when short <= target_size, else aspect-preserving downsample + center-crop. Boundary at short == target_size deliberately skips the crop (docstring and test test_short_side_at_target_passes_through_without_cropping agree).
• Test coverage adequate for the new branches (square smaller, non-square smaller, shorter-side-at-target non-square); existing landscape/portrait/already-square cases still exercise the resize+crop path.
• No downstream code in annotate_subtasks.py assumes square output — frames flow through _to_jpeg_bytes/_to_b64_jpeg to the API, both PIL JPEG and the Anthropic/Gemini image inputs accept arbitrary aspect ratios.
• Module docstring, function docstrings, and --target-size CLI help kept in sync with the new behavior.
• One mild caveat (not a defect): for an input whose shorter side is exactly target_size but is non-square (e.g. 448×600), the new behavior keeps the frame at 448×600 instead of cropping to 448×448, so for that specific shape image-token cost rises slightly vs. before. Common video inputs (≥VGA, where short > target_size) are unaffected, and frames with short < target_size clearly win — overall the PR's stated goal holds.