Two-step pipeline: task_allocate("kiwi", "recipe_scan", service_hint="cf-docuvision")
acquires a docuvision allocation, calls /extract per image to get OCR text, then
LLMRouter structures the combined OCR output into recipe JSON via the text
extraction prompt.
Also fixes DocuvisionClient bugs:
- POST field was "image" (ignored by Pydantic) — should be "image_b64"
- Response read "text" key — docuvision returns "raw_text"
- Add hint parameter (use "text" for recipe cards, dense prose)
- Configurable timeout (default 120s; docuvision lazy-loads model on first request)
_parse_json_from_text always returns a dict (never None), so the
previous `if parsed is not None` guard was permanently true — garbled
docuvision output would return an empty skeleton instead of falling
through to the local VLM. Replace the check with a meaningful-content
test (items or merchant present). Add two tests: one that asserts the
fallthrough behavior on an empty parse, one that confirms the fast path
is taken when parsing succeeds.
Introduces a thin HTTP client for the cf-docuvision service and wires it
as a fast path in VisionLanguageOCR.extract_receipt_data(). When CF_ORCH_URL
is set, the pipeline attempts docuvision allocation via CFOrchClient before
loading the heavy local VLM; falls back gracefully if unavailable.
