[FEAT] Prompt Evaluation Framework with Acceptance Metrics

[Anishp-cell]

Now that initial prompt unit tests have been merged (#13),
there is an opportunity to improve the evaluation framework by:
• Introducing acceptance metrics (e.g. ≥70% correct extraction across N receipts)
• Classifying failure types (OCR error vs hallucination vs missing field)
• Decoupling test execution from app packaging more cleanly
• Making prompt evaluation reproducible & comparable across model versions
This would move prompt testing from “smoke tests” to a measurable evaluation framework.
Happy to take this on if aligned with the roadmap.