fix: derive eval subjects from document content (prof-cgu)

[ckrough]

Why: Encode the synthetic-eval subject convention in code so ground-truth rows describe document content rather than the document-type form the file's header prohibits.

Changes:

• Derive each row's subject via a model call against generated text, with a structural forbidden= guard that rejects the f"{category} {doctype}" form
• Add scripts/backfill_eval_subjects.py to apply the same derivation in place across existing PDFs, preserving comments, key order, and entity fields
• Validate --ai-model against MODEL_PRICING at startup; collect-then-apply on cost-cap abort; share an atomic-write helper with the generator
• Add tests for sanitizer normalization, forbidden-form retry, empty-text raise, and fallback behavior
• Rewrite the 80 subjects in eval/ground_truth/synthetic.jsonl to content-based form

Issues: prof-cgu