docs(readme): HR2 reviewer entry (purpose + setup paths)

[ChunkyTortoise]

What

README-only CONT-HR2 reviewer entry: a plain-language purpose before the score wall, and three distinct setup paths.

Why

Reviewers currently meet the 95.5% replay before the product purpose, and the reviewer steps mix offline replay, the fixture UI, and inspection without a separate full-stack path.

Changes

• HR2-1: Purpose statement before the first evaluation section. It reuses capabilities already in "What this does": upload PDFs and images, classify with cost-aware routing, extract structured fields with the two-pass Claude pipeline, store embeddings in pgvector, and search with agentic RAG. The existing "What this does" section stays, with a short cross-reference.
• HR2-2: Reviewer path is three blocks:
  1. Offline fixture replay (no API key), from the repository root: python scripts/eval_offline_replay.py --floor 0.85.
  2. Fixture-backed UI demo (no API key): DEMO_MODE=true streamlit run frontend/app.py, matching DEMO.md and frontend/app.py.
  3. Full configured services: link to Install (cp .env.example .env, ANTHROPIC_API_KEY, GEMINI_API_KEY, docker compose up -d). No new install recipe.

Frozen claims preserved

• 28 committed prediction fixtures
• 0.9555 / 95.5% weighted field-level accuracy (not F1, not live-model)
• 72 lookup cases; 44 pending
• 200 non-metadata authoring cases / 202 JSONL lines (two metadata rows)
• 80% CI coverage gate
• Retrieval recall, support, and abstention unmeasured
• No new merge-blocking claim (CI checks still described as not enforcing merge protection)
• No new paid or independent labeling beyond wording already on main
• docs/retrieval-extraction-evidence.md is untouched

Testing

• git diff --check (whitespace)
• git diff --name-only origin/main is only README.md
• Frozen metrics above still present with the same semantics
• git diff origin/main -- docs/retrieval-extraction-evidence.md is empty
• Every scripts/*.py named in README exists on this branch (eval_offline_replay.py, eval_gate.py, benchmark.py, run_eval_ci.py)
• pytest tests/ -v, eval gate, mypy, ruff: not run; no app, test, or workflow changes

Screenshots

None. README only.

How to verify

git diff --check origin/main
git diff --name-only origin/main
git diff origin/main -- docs/retrieval-extraction-evidence.md
python scripts/eval_offline_replay.py --floor 0.85

Confirm the opening paragraph states upload, classify, extract, and search before ## Deterministic eval replay, and that the reviewer path lists the three setups above.

  

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-23T02:01:09.843040Z	cf0874c	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.