feat(ocr): per-sender model registry and /train-sender endpoint

engines/kraken.py:
- Add _SenderModelRegistry with LRU eviction (max configurable via
  OCR_MAX_CACHED_MODELS env var), double-checked locking, invalidate(),
  and path whitelist (/app/models/ only)
- Add _load_sender_model() helper for testability
- extract_page_blocks() and extract_region_text() accept optional
  sender_model_path; route to sender registry when provided

models.py:
- OcrRequest gains senderModelPath: str | None = None field

main.py:
- /ocr and /ocr/stream pass request.senderModelPath to Kraken engine
- New /train-sender endpoint: validates output_model_path, runs ketos
  train with base model as starting point, invalidates sender cache

docker-compose.yml:
- Add OCR_MAX_CACHED_MODELS: "5" to ocr-service environment

test_sender_registry.py:
- 4 tests: cache hit, LRU eviction, invalidate, path traversal guard

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docker-compose.yml

@@ -97,6 +97,7 @@ services:
       DETECTOR_BATCH_SIZE: "8"
       OCR_CLAHE_CLIP_LIMIT: "2.0"   # CLAHE contrast limit (multiplier of average histogram frequency)
       OCR_CLAHE_TILE_SIZE: "8"      # CLAHE tile grid size (NxN tiles per page)
+      OCR_MAX_CACHED_MODELS: "5"    # LRU cache size for per-sender Kraken models
     networks:
       - archive-net
     healthcheck: