Ler documento de veículo (DUA/livrete) e devolver campos para autofill

OCR a vehicle document (JPEG/PNG/PDF) via Claude Vision and return RAW TEXT fields for the create-vehicle form to prefill (id resolution happens in the FE, section-scoped — never here).

Accepts ONE OR MORE images so the caller can send the frente AND the verso — the technical fields (cilindrada, potência, combustível, cor, lugares) live on the verso of the DUA/livrete, so a front-only photo can't read them. All images are read together in a single vision call and billed as ONE read.

Unreadable/unsupported input returns success=False (never a 5xx); billing only fires on success. Heavy sync work (CV + PyMuPDF + Anthropic + pyodbc) runs in the threadpool so the event loop is never blocked (same pattern as /invoices/process-qr).