Lock recovery to active solution file and harden validation commands

[mmprotest]

Motivation

• Prevent recovery-mode escape by pivoting to a different top-level entrypoint file and stop delete-and-regenerate tactics.
• Make validation/launch steps conservative so shell constructs cannot trivially mask failures.
• Keep changes minimal and localized to recovery, tooling, and validation handling without large redesigns.

Description

• Add a compact runner state field ._active_solution_file and initialize/reset it in Runner (villani_code/state.py) to track the active solution path.
• Seed ._active_solution_file when a focused mutation/verification target is first registered in execute_tool_with_policy (villani_code/state_tooling.py) and from run_verification when appropriate.
• Anchor recovery protections in small_model_tool_guard (villani_code/state_runtime.py) to the active_solution_file (falling back to failing_file), blocking deletion/full-rewrite of the active solution path while allowing edits to helper modules and returning compact JSON feedback when blocked.
• Detect suspicious shell masking patterns via _command_has_masking_patterns, classify validation/launch intent with _is_validation_or_launch_command, and extract python entrypoint targets with _extract_command_python_target in state_runtime.py, blocking obvious masked validation/launch commands and preventing run/validate pivots to a different entrypoint during recovery.
• Exclude masked command records from validation artifacts in collect_validation_artifacts (villani_code/state_execution.py) so masked commands cannot advance verification.
• Preserve conservative failing/active alignment so helper-module failures do not automatically switch active_solution_file but hard failures on the active file keep both fields aligned; events include active_solution_file where relevant.

Testing

• Added unit tests and changes in tests/test_state_tooling_mutation_policy.py, tests/test_state_runtime.py, and new tests/test_state_execution.py covering active-solution locking, allowed helper edits, parallel-entrypoint blocking, and masked-command handling.
• Run: pytest -q tests/test_state_tooling_mutation_policy.py tests/test_state_runtime.py tests/test_state_execution.py produced 53 passed (all targeted tests passed).
• No unrelated failures observed in this targeted test run.

---

Codex Task