Add event-driven patch-recovery controller for failed Patch/Edit operations

[mmprotest]

Motivation

• Improve edit-loop reliability by handling failed patch/edit tool results with a small, generic, event-driven recovery controller so the model is steered to produce a smaller patched retry instead of drifting to immediate Bash/Python file rewrites.

Description

• Add operational recovery state to the runner and mission state via runner._patch_recovery_state, runner._patch_recovery_pending_turn, runner._patch_recovery_attempts_by_file and mission fields last_patch_failure, recent_tool_failures, and dirty_temp_files (files: villani_code/state.py, villani_code/mission_state.py).
• Implement patch-failure capture and bounded file-context refresh in state_runtime: extract target file and failed hunk, match anchor lines in the current file, build a compact context slice (default ~20 lines radius, hard cap ~120 lines), record structured state (target_file, error, failed_hunk, refreshed_context, attempts_for_file, next_required_action) and emit concise events (patch_failure_detected, patch_recovery_context_refreshed, patch_recovery_instruction_injected) (file: villani_code/state_runtime.py).
• Inject a one-turn, strongly-worded recovery instruction into the next model turn via prepare_messages_for_model, and clear recovery state when a subsequent successful Patch/Edit targets the same file (files: villani_code/state_runtime.py, villani_code/state.py).
• Add a lightweight, generic Bash/Python guard in state_tooling that blocks or discourages immediate shell-based rewrites of the same target file while patch recovery is active and fewer than two refreshed patch retries have run, while still allowing read-only and test/inspection shell commands (file: villani_code/state_tooling.py).
• Add focused unit tests covering structured failure recording, refreshed context extraction, one-turn instruction injection, bash-guard blocking behavior, read-only/test allowlist, and non-triggering on successful patches (file: tests/test_patch_recovery_controller.py).

Testing

• Ran pytest -q tests/test_patch_recovery_controller.py tests/test_state_tooling_mutation_policy.py and observed all tests passing (12 passed).
• Ran pytest -q tests/test_loop.py::test_tool_result_followup_is_pure_tool_result_message tests/test_villani_failure_modes.py::test_first_attempt_locked_target_rejects_extra_file_before_verification and observed both tests passing (2 passed).

---

Codex Task