fix(detect): Filter bright rectangles from TOC leak bbox matching

[mlissner]

Summary

Fix false TOC leak detections on PDFs with colored highlight annotations (pink, yellow, gray highlights over text in academic papers, etc.).

Root cause: get_good_rectangles() returns all opaque filled rectangles, including colored highlights. Their bboxes were being added to redaction_bboxes before any color filtering, polluting the TOC leak matcher.

Fix: Filter rectangle bboxes by _is_dark_color on fill color before adding to redaction_bboxes. This excludes bright highlights while keeping dark rectangles (including applied redactions with no text underneath).

Note: The initial suggestion was to move bbox collection after the pixmap filter, but that would have broken TOC leak detection for properly applied redactions (text removed, black bar left) since those have no text and don't survive the pixmap pipeline.

Test plan

• New test test_highlight_annotations_toc_no_results with real academic paper PDF
• Existing test_toc_leak (7-page synthetic) still passes — applied redaction on page 1 still detected
• Existing test_toc_leak_real (JOSH MERRITT) still passes
• All 42 tests pass
• mypy and pre-commit pass
• Merge after feat(detect): Detect SSNs hidden under white rectangles #217 lands

🤖 Generated with Claude Code

[claude]

LGTM — targeted fix for false TOC leak detections from colored highlight rectangles, with clear test coverage.

Extended reasoning...

Overview

This PR fixes a false-positive bug in TOC bookmark leak detection: get_bad_redactions was adding bboxes of all opaque rectangles (including colored highlights used as markup in academic papers) to the redaction_bboxes list used by get_toc_leaks, causing bookmarks near highlighted (non-redacted) text to be wrongly flagged as leaking redacted content. The fix filters these bboxes to only dark-colored rectangles via the existing _is_dark_color helper, before they are added to redaction_bboxes. It deliberately leaves the unfiltered good_rectangles list untouched for the actual text-under-rectangle detection path, and preserves TOC-matching for genuinely applied redactions (black bar, no text) by checking fill color directly rather than gating on the pixmap filter.

Security risks

None. This is pure PDF-content heuristic logic with no user input handling, network calls, or auth/permissions surface.

Level of scrutiny

Low-to-moderate. The change is a small, localized addition (an if filtering which bboxes get appended) inside a well-tested detection pipeline. The two candidate concerns raised by the automated bug hunt (bbox/redaction desync, dropping legit applied-redaction bars) were both examined and are addressed by the PR's design rationale (dark rectangles, including bar-only applied redactions with no text, are still included) — I independently confirmed rect.fill is already normalized RGB elsewhere in this file (e.g. get_dark_highlight_annotations), so the int(c * 255) conversion is consistent with existing usage.

Other factors

New test test_highlight_annotations_toc_no_results uses a real academic-paper PDF per the project's testing rules, and the full test suite (42 passed, 1 expected-failure) passes locally, including all TOC-leak and highlight-related tests. CHANGES.md is updated per project convention. No outstanding review comments in the PR timeline.