[charan-rathore]

Summary

Fixes #35. ocr_region joined a low window to the menu bar at the top of the display, then OCRed the entire rectangle between them. On a window over another app, unrelated background text entered the classifier's screen read and made OCR expensive. The crop now stays within the frontmost window plus its existing margin, clamped to the display. With no usable window, it keeps the full-display fallback.

This deliberately uses one crop rather than changing the OCR cache to manage two regions. The reporter measured 46.4 seconds for the joined region and 0.33 seconds for the window-only region on their 5K Finder repro. Those are reporter measurements, not a live replay by this PR.

Checks

• uv run --no-sync ruff check .: passed.
• uv run --no-sync ruff format --check .: passed.
• uv run --no-sync pytest -q: 612 passed, 3 skipped, 1 xfailed on Linux.
• Fake-screen tests cover the low-window crop and OCR box offsets, a window at the top, offscreen edges, no window, and a degenerate window. Existing AX tests show that labelled menu-bar controls remain in the accessibility walk. PR macos: walk the focused window and menu bar #42 also preserves the menu-bar root when it scopes the macOS walk to the front window.

Caveat

Text visible only in the menu bar that is not available through accessibility will no longer be OCRed. There was no live Finder/OCR replay here; the focused tests use fake captures and a fake OCR reader.

Instinct assisted.

[charan-rathore]

Closing this as superseded. Main now deliberately keeps the menu bar strip in the OCR region for a focused window (see the ocr_region tests in tests/test_ocr_cache.py on main), and this PR removes that strip, so it would work against the current design. Sorry for the noise.