Remove Phase-1 hard completion audit gate (preserve tool-call recovery & shell guardrails)

[mmprotest]

Motivation

• Phase 1 introduced an end-of-run structural completion audit that is too heavy and causes verification loops and false-negative completion blocking, so it must be removed while retaining other Phase 1 hardening that improves execution discipline.

Description

• Deleted the villani_code/completion_audit.py module which implemented claim-to-path extraction, structural checks, CompletionAuditResult, repair-brief generation, and completion-blocking logic.
• Removed the run_completion_audit(...) import and all calls to run_completion_audit(...) in both finalization paths inside Runner.run in villani_code/state.py, and removed emission of completion_audit_blocked events and repair-brief message injection that looped the agent back into repair mode.
• Kept and did not alter the malformed structured tool-call recovery code path and its eventing (tool_schema_recovery_triggered) in villani_code/state.py so repeated malformed tool calls still trigger schema-focused recovery.
• Preserved environment-aware shell guardrails in villani_code/shells.py and the related command_rejected debug/event visibility in state.py (Windows-incompatible bash forms remain rejected).
• Updated tests: replaced the completion-audit-specific blocking test in tests/test_runner_phase1_hardening.py with a test asserting completion is not blocked after a text-only assistant turn, while keeping tests that verify malformed tool-call recovery and shell guardrails.

Testing

• Ran pytest -q tests/test_runner_phase1_hardening.py and the suite passed (3 passed).

---

Codex Task