Use ExecutionMemento as regular-runner carry-forward state

[mmprotest]

Motivation

• Reduce reliance on a long accumulated transcript by carrying a compact structured execution-state snapshot between regular runner turns.
• Provide a machine- and model-readable memento that states: objective, belief, evidence, scope, attempts, verification outcomes, and next action.
• Make the change minimally invasive, preserving existing helpers and fallback behavior when the memento is not available.

Description

• Added villani_code/execution_memento.py implementing ExecutionMemento dataclass and helpers: build_execution_memento, save_execution_memento, load_execution_memento, render_execution_memento_for_model, and build_local_evidence_block and persisted artifacts to .villani_code/missions/<mission_id>/execution_memento.json (/optional .md).
• Extended MissionState with last_memento_path: str and last_memento_turn_index: int and updated serialization/roundtrip in villani_code/mission_state.py.
• Triggered memento creation only at meaningful boundaries by adding _refresh_execution_memento() and invoking it from Runner._update_mission_state when key fields change and from validation/result code paths in state_runtime.py.
• Changed regular-runner prompt assembly to be memento-first: _inject_projected_context now prefers the rendered EXECUTION STATE + tiny LOCAL EVIDENCE block and falls back to the legacy Mission context packet: when required memento fields are missing; prepare_messages_for_model trims long history aggressively for regular (non-small, non-autonomous) runners and preserves tool_use/tool_result adjacency when keeping a short tail.
• Kept small-model behavior and existing context_projection helpers intact for backward compatibility and preserved tool sequence integrity.

Testing

• Ran the focused test subset with pytest -q tests/test_mission_state_runtime.py tests/test_state_runtime.py tests/test_session_compaction.py and all tests passed (57 passed).
• Added/updated targeted tests covering memento persistence, mission-state linkage, memento-first injection and fallback, non-duplication of injection, transcript tail trimming and tool sequence integrity (files updated: tests/test_mission_state_runtime.py, tests/test_state_runtime.py, tests/test_session_compaction.py).

---

Codex Task