feat(degradations): model capture conditions, moving boxes with the pixels

[hBouanane]

6 of 9 in the v1.0.0 series. Base: feat/document-engine.

The design point

A degradation takes the annotation as well as the image, and returns both.

A rotation that moves ink 8° but leaves the boxes where they were produces a dataset that looks fine and teaches a detector to be systematically wrong. That failure is invisible until training underperforms for no obvious reason.

What

• base.py — Degradation protocol, DegradationPipeline with per-degradation probabilities and a record of exactly what fired, and map_page(), which rebuilds a whole annotation through an arbitrary point mapping.
• geometric.py — Rotation and PerspectiveWarp, each deriving an explicit forward homography and pushing every box, word and table cell through it. Rotated words gain polygons, since an axis-aligned box no longer describes them tightly.
• photometric.py — sixteen degradations chosen for physical cause rather than visual effect: ink spread/erosion, bleed-through from the reverse of a thin sheet, downscaling (the failure mode that destroys Arabic diacritics and that blur does not reproduce), JPEG artefacts, shadow, vignette, glare, stains, folds.
• presets.py — clean / scan / photo / fax / archive, ordered along the physical capture chain (optics → shading → sampling → compression), so a corpus is composed by capture condition rather than by an undifferentiated cloud of noise.

Verified

365 tests green (79 new). The central test measures the actual ink box in the degraded image and requires the annotation to track it at IoU > 0.8, rather than merely checking that something changed.