Maya's extractions use a tuned prompt

[danielnaab]

User Story

As Maya, in order to get better extraction quality without changing the underlying model, I want an extraction variant whose system prompt has been automatically optimized against the evaluation suite.

Preconditions

• Story 10: Maya chooses how her form was extracted #58 (Story 10 variant picker) merged to main

Acceptance Criteria

• New service src/services/prompt-optimization/ — thin wrapper around TextGrad (or equivalent) that drives our existing EvaluationKind harness as its training signal
• Optimization script produces a concrete optimized prompt for extraction, committed as a file under src/services/extraction/prompts/optimized-v1.txt (or similar)
• New variant extraction/sonnet-optimized-v1 that loads the optimized prompt at construction
• Extraction tab in Settings → Variants lists the optimized variant
• Evaluation run comparing baseline sonnet vs sonnet-optimized-v1 on all fixtures
• New catalog page catalog/experiments/pdf-field-extraction/sonnet-optimized-v1.md including: optimization setup (epochs, batch size, forward/backward models), before/after prompt snippets, and metric deltas
• catalog/experiments/_roadmap.md updated with shipped status and one-line finding
• Harness documented well enough that follow-on stories can run it against shaping and filling suites

Success Metrics

• Optimized variant beats baseline on at least one metric (recall, precision, or sensitivity)
• Optimization is reproducible via a single CLI command

Notes

• Class topic: prompt optimization, Assignment 10
• Use Opus as the backward model if it fits the budget; Sonnet otherwise
• Document hyperparameters for reproducibility
• Follow-on stories (not in scope here): shaping/sonnet-optimized-v1, filling/sonnet-optimized-v1 — same harness, different eval kind

Definition of Done

• Acceptance criteria met
• Tests pass
• Type checking passes
• CI pipeline green
• Deployed and demoable