Reduce end-to-end runtime while keeping the system fully offline and CPU-only.
Acceptance criteria:
- Profile preprocessing, full-image OCR, local OCR crops, and assignment stages separately.
- Cache reusable OCR/preprocessing results per request where safe.
- Avoid full-page rotated OCR except as an explicit fallback.
- Document before/after runtime on representative benchmark images.
Reduce end-to-end runtime while keeping the system fully offline and CPU-only.
Acceptance criteria: