Release note: client OCR fallback ladder

What changed

The client scan now loads OCR engines through an ordered ladder (`lib/ocr/engine-ladder.ts`) instead of a single engine:

1. MRZ model from the browser cache (`public/tessdata/mrz.traineddata`, Doubango `tessdata_fast` int model, BSD-3-Clause).
2. MRZ model re-downloaded with `cacheMethod: "refresh"` if the cache attempt failed (stale or corrupt model).
3. Stock `eng` data from the CDN — last attempt only.

A shared singleton worker (`lib/ocr/shared-worker.ts`) applies the same OCR parameters to whichever engine loads, and `activeOcrEngine()` reports which one won.

Behaviour guarantees

• OCR data tests (stock `eng`) are used only when everything falls — never first, never by default.
• Functionality is maintained on the fallback: band read, candidate ranking, look-alike repairs, and send-to-Validate/Compare all work identically.
• Failed loads are never cached; every read retries the ladder from the MRZ model.
• The user is told which engine read the document and can trigger "Try the MRZ model again" at any time.

Validation

`lib/ocr/engine-ladder.test.ts` covers the ladder: first success, fallback after MRZ failure, last-error rethrow, attempt reporting, and empty-ladder rejection. All tests, `tsc --noEmit` and ESLint must stay clean.