[Moskize91]

Later OCR passes can return a page-sized box for a footer digit even though most of the input has been erased. Clip each later-stage layout to the largest rectangle inside its original box that avoids every previously erased pixel; discard fully covered layouts. Accumulate actual masks across arbitrary stage counts, restore resized coordinates before clipping, and rebuild structured output after correction.

Exports clip_layouts_to_unmasked_area for consumers that orchestrate their own OCR calls. It accepts the same inclusive mask coordinates as Pillow redaction; layout bounds remain half-open. Changed boxes discard stale polygons while retaining text and raw diagnostics. Equal-area results prefer top, then left. Coordinate compression avoids scanning page pixels.

The captured DeepSeek-OCR page-4 response had box (0,0,2474,3497) for 4. Its actual masks produce (0,3158,2474,3497), which retains the complete real digit at (1228,3247,1249,3278) without overlapping erased pixels. This is a fallback for invalid OCR geometry and may discard text in smaller remnants.

Bumps both package version declarations from 1.2.1 to 1.2.2 and documents cumulative masks and custom-pipeline integration. pdf-craft's separately orchestrated asynchronous OCR path must explicitly adopt the helper in a subsequent change; this PR does not change that repository or deploy services.

Validation: all 36 lightweight tests pass, including captured geometry, inclusive boundaries, four-stage cumulative masking, image rescaling, per-call isolation, structured output, and 150 seeded cases compared with exhaustive rectangle search (including reversed mask order). Pylint: 10.00/10; git diff --check passes. No CUDA or paid OCR calls required.