Add lightweight runner guards for repeated failure thrashing

[mmprotest]

Motivation

• Reduce obvious thrashing where the agent repeatedly issues the same failing mutation or brittle shell form without adding determinism or a heavy recovery subsystem.
• Keep interventions minimal, visible in existing events/transcripts, and reversible after one simple grounding step.
• Preserve exploratory behavior and avoid task-specific heuristics or planner-style logic.

Description

• Added a tiny rolling failure memory and transient guard state on Runner: _recent_failed_actions, _recent_failed_shell_forms, _surface_grounding_block, _shell_retry_block_form, and related pending message fields to hold short-window failure history. (files: villani_code/state.py)
• Implemented compact helpers for normalization/classification: _normalize_failure_reason, _failure_kind, _shell_failure_from_result, _normalize_shell_command, _is_brittle_shell_command, _is_grounding_action, _record_failure_and_update_guards, and _check_transient_action_guards, enforcing the exact trigger rules in the last-3/last-5 windows. (file: villani_code/state.py)
• Injected a single short recovery message into the next tool-result turn when triggers fire and require one grounding action (read/inspect or non-mutating shell) before allowing the same mutation surface again; the pending message is cleared after being attached to the next user tool-result payload. (file: villani_code/state.py)
• Added a lightweight pre-policy guard hook in execute_tool_with_policy to detect and block transient guarded actions (including brittle near-duplicate shell retries), emit a transient_action_guard_blocked event, and return an explicit short block message as a normal tool error. (file: villani_code/state_tooling.py)
• Added focused tests exercising the new guards and clearing behavior in tests/test_loop.py. (file: tests/test_loop.py)

Testing

• Added unit tests in tests/test_loop.py covering: repeated tool failures triggering a regrounding intervention and temporary mutation block, grounding clearing the block, near-duplicate brittle shell retry blocking and clearing, and that isolated single failures do not trigger guards.
• Ran the focused test selection with: pytest -q tests/test_loop.py -k 'failure_loop_guard_blocks_mutation_until_grounding or brittle_near_duplicate_shell_retry_guard_and_clear or single_failures_do_not_trigger_thrash_guards', which passed with 3 passed, 25 deselected.
• All modified tests in the focused run succeeded.

---

Codex Task