feat(agent): six-step agent flow with Streamlit Flow page and group guide

[jsshin2022]

Stacked on #1 — merge #1 first, then this PR retargets to main.

What's added

• Six-step agent flow (core/agent/), mirroring the track's pipeline (collect → structure metrics → estimate → report, automated by an agent):

Step	What it does	Week
1. Define problem	Validate plan.yaml; pre-registration gate (plan must be committed before estimation)	W3
2. Collect	Run the data adapter; record source, license, snapshot hash	W2
3. Structure metrics	Column/type checks, missing rates, pre/post and treated/control counts	W3
4. Estimate	run_plan (DiD / event study / ITS) + placebo test; the LLM never computes numbers	W4–5
5. Guard	Abstention verdict + overclaim linter on report text (e.g. no "no effect" when the CI crosses zero)	W6
6. Report	report.md, figures, run_manifest.json (git SHA, versions, data/plan hashes)	W7

• Runs offline with no API key (pure-Python runner) or via LangGraph (engine=langgraph). Optional LLM narrative through any OpenAI-compatible endpoint (OPENAI_BASE_URL, OPENAI_API_KEY, LLM_MODEL); its text is linted before use.
• CLI: make flow CASE=cases/_example_night_clinic
• Streamlit Flow page (app/pages/1_Flow.py): run or replay a flow; per-step status, quality table, estimate and verdict, guard findings, report preview. Shows a SYNTHETIC banner for synthetic cases.
• Sample outputs committed for cases/_example_night_clinic (flow_log.json, run_manifest.json, report, figures) — synthetic data, not a real policy effect.
• docs/ops/group-guide.md: group roles mapped to the six steps, weekly milestones, branch/PR rules.
• CI now also runs the example flow offline and uploads the log and report as artifacts.

Fixes

• Flow logs store paths relative to the case folder (no local machine paths in a public repo).
• The example-case test runs in a temp copy, so tests no longer modify tracked files.

Verification

• pytest: 36 passed; ruff check / ruff format --check: clean
• Example flow: event study −5.03 (95% CI −6.11 to −3.96) vs. true effect −5.0 → verdict identified; overall status warn because the data is synthetic.

Remaining

• Callaway–Sant'Anna estimator for staggered adoption (most real cases); currently flagged and withheld.
• Adapters for data.go.kr / Seoul Open Data; KOSIS adapter still unverified against the live API.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EMnmZPBD5PQJW1UQAARN7q

---

Generated by Claude Code