rename package doc_agent -> docfield, distribution -> docfield-extract

eval: add --retry-errors to re-predict only the entries that failed

web: trim docstrings and drop the copy-preference test

web: explain the decision in the output, in plain language

validation: monetary tolerance scales with rounded terms, not with value

eval: report documents that never reached the model, grouped by cause

validation: S2 skips an absent currency instead of failing it

eval: add tuning/held-out split selection to the score command

eval: score under current rules with --revalidate, and detect cache drift

phase 5: set CONFIDENCE_THRESHOLD=0.50, record SROIE results (T10)

phase 5: evaluation harness (SROIE)

fix(lint): clear 5 ruff F401/F841 errors found by verifier

feat(watcher): folder watcher + batch runner (T8)

feat(store): SQLite persistence + CSV export (T7)

feat(gemini): Gemini backend + vision-direct image acquire (T2)

phase 3.2: content-hash idempotency helper

phase 3.1: core pipeline orchestration (stub-backed smoke test)

phase 2.4: backend interface + factory + offline stub backend

phase 2.1: modality detection

phase 1.3: confidence scoring + routing decision

phase 1.2: validation rules (hard/soft + arithmetic checks)

phase 1.1: pydantic document schema + normalizers

phase 0.2: config loader with validation

phase 0.1: project scaffold (uv, py3.11, package layout)