Add lightweight per-tool malformed invocation recovery

[mmprotest]

Motivation

• Prevent model/runner loops where the model repeatedly emits malformed tool invocations for the same tool and the run spirals without useful guidance.
• Keep the fix small, local, tool-agnostic and nondeterministic so normal exploratory behavior and existing runner control flow remain unchanged.

Description

• Add a compact per-tool recent-failure tracker and transient degradation state on Runner using a short sliding window with _MALFORMED_TOOL_WINDOW = 4 and _MALFORMED_TOOL_THRESHOLD = 2 and lightweight escalation levels.
• Detect only malformed invocation failures (schema/validation style errors returned by execute_tool) via _is_malformed_tool_invocation_failure, and do not count ordinary execution failures.
• When repeated malformed invocations are observed, replace the verbose schema error payload with a very short recovery nudge produced by _recovery_message_for_tool, asking the model to re-ground before retrying while leaving the model free to choose how to recover.
• Make the degraded state transient and inspectable: failures are recorded in a per-tool deque, degradation levels and recovery turns are tracked, and _clear_tool_degradation clears state after a valid invocation.
• Integrate recovery with a single, surgical insertion point after tool execution (result = self._apply_malformed_tool_recovery(...)) so the runner behavior and flow are preserved without adding planners or a fallback mode.
• Add focused unit tests that simulate sequences of tool uses and assert the recovery behaviour is triggered, cleared, and escalated correctly.

Testing

• Added tests to tests/test_loop.py covering: test_no_recovery_message_on_single_malformed_invocation, test_recovery_triggers_after_repeated_malformed_invocations, test_degraded_state_clears_after_valid_invocation, test_immediate_relapse_after_recovery_escalates, and test_ordinary_execution_failures_do_not_count_as_malformed and supporting test helper SequencedToolClient.
• Ran the new/modified test file with pytest -q tests/test_loop.py and the file passed: 30 passed.
• The targeted recovery tests (the added cases) were executed and passed.

---

Codex Task