Why

Suspicious visual overlaps need a diagnostic signal so generated specs can be repaired before rendering, without flagging compositions that only overlap in bounding boxes or transparent pixels.

Changes

• Add layer-overlap diagnostics from measured layer geometry and render order.
• Use bounding-box overlap as a broad-phase filter, then measure actual visible alpha-mask intersection for text, images, SVGs, and shape layers.
• Report both bbox_overlap_pct and visible_overlap_pct with overlap pixel counts and a concrete move suggestion.
• Preserve a fast path for simple opaque rectangles.
• Allow text fully contained on a lower backdrop while flagging text/text, substantial partial overlaps, complete non-backdrop coverage, and non-text layers covering text.
• Recompute per-pass alpha masks so repeated diagnostics reflect the current canvas layers.
• Cover text/text, text/backdrop, partial overlap, ellipse corner false positives, transparent PNGs, glyph holes, full coverage, repeated diagnostics after layer mutation, non-overlap, touching edges, invisible layers, and grouped layers.

Validation

• uv run ruff check quickthumb/_diagnostics.py tests/test_diagnostics.py
• uv run pytest tests/test_diagnostics.py -q passes: 47 tests.
• uv run pytest reports 703 passed, 24 skipped, 52 inline-snapshot teardown errors in rendering snapshot tests, and one external httpbin.org 503 failure in test_snapshot_url_image_background_basic.