refactor(policy): extract combat selection and context modules

docs/POLICY.md

@@ -0,0 +1,66 @@
+# Policy structure
+
+## Current boundaries
+
+```text
+run.py                 observe, execute, report results, record the session
+brain.py               entry point, dispatch, navigation, shops, minigames
+policy/
+    __init__.py        package marker; no registration or initialization
+    context.py         Decision, PendingAction, PolicyContext
+    combat.py          combat proposals and model questions
+    selection.py       card, relic, bundle, hand, and reward proposals
+facts.py               observation parsing and combat calculations
+```
+
+`brain.decide()` remains the policy entry point. The runner owns one
+`PolicyContext` per session and reports action results to it.
+`brain.Decision` and `brain.PolicyContext` remain available as direct imports
+of the shared types, so the runner interface does not change.
+
+Policy modules may ask Jev for preferences. They must not call the game API.
+`context.py` does not call either service. Combat arithmetic remains in
+`facts.py`, not duplicated inside combat policy.
+
+The dependency direction is simple:
+
+- The runner uses `brain`.
+- `brain` uses the combat, selection, and context modules.
+- Combat and selection use context, facts, and the model client.
+- Context uses only the Python standard library.
+- No policy module imports `brain` or the runner.
+
+## State and configuration
+
+A proposal does not update execution memory. Accepted requests are reconciled
+with fresh observations. See [policy-state behavior and limits](research/13-policy-state.md)
+for the evidence required by each guarded action.
+
+The existing card-skip setting stays in `brain.CARD_SKIP_POLICY`. The dispatcher
+passes its value explicitly to `selection.card_reward_decision(skip_policy=...)`.
+This avoids a circular import and preserves the existing command-line behavior.
+It is configuration, not per-screen action memory.
+
+## Why stop here
+
+These are plain modules, not a plugin framework or class hierarchy. Navigation,
+shops, and minigames stay together until a further split helps development.
+
+The extraction does not change game decisions, confidence gates, fallbacks,
+state reconciliation, or recording. Run identity and persistent-deck provenance
+remain separate work.
+
+## Extraction verification
+
+- 262 existing assertions passed. No test files or assertions were added for the extraction.
+- Thirteen isolated whole-process scenarios passed with local HTTP fixtures.
+- All 346 stored observations produced identical decisions before and after extraction,
+  in both model-free mode and local-model-stub mode: 692 comparisons.
+- The local model requests also matched, including their state and questions.
+- Thirty-six function/class definitions retained identical abstract syntax trees.
+  Only the two dispatch functions and explicit card-skip parameter changed.
+- Dataset integrity, Python compilation, and shell syntax checks passed.
+
+The differential and whole-process probes were temporary. They made no live
+game or paid model calls. These checks establish refactoring equivalence for
+the tested inputs, not correctness for every possible game state.