Harden regular runner against context loss and false completion

[mmprotest]

Motivation

• The regular-runner can lose durable task state in long conversations so the model emits generic conversational replies and the runner incorrectly treats them as completion.
• Existing execution-memento support was insufficient when saved mementos are missing or incomplete, allowing prompt trimming to drop mission-critical fields.

Description

• Add build_fallback_execution_state_block(runner) to villani_code/execution_memento.py to produce a compact, structured fallback block containing Objective, Success, Current verification, Next action, and optional Constraints.
• Update _inject_execution_state_memory in villani_code/state_runtime.py to prefer a valid saved memento but always inject the fallback block when memento data is missing or invalid, so regular turns never lack durable task state.
• Strengthen _trim_regular_turn_messages to preserve all system messages, an earliest non-tool-result user task anchor, and the recent atomic tail while keeping tool-use/tool-result pairs intact.
• Add a final invariant _ensure_regular_execution_state_invariant in prepare_messages_for_model that guarantees the prepared prompt contains Objective:, Success:, and Next action: and injects the fallback block into the latest safe user message (or appends a user message if needed).
• Add a minimal completion guard in the main loop (villani_code/state.py) that detects obvious generic no-tool fallback replies when verification is unresolved and prevents them from being treated as successful completion, requesting a context-restore turn and terminating with stalled_context_loss after repeat hits.

Testing

• Added and updated focused tests in tests/test_state_runtime.py covering fallback injection when the memento is missing or incomplete, trimming behavior that preserves the task anchor and tool-sequence integrity, avoidance of duplicate execution-state stacking, and preference for valid saved mementos.
• Added a regression test in tests/test_bounded_execution.py to assert a generic no-tool assistant reply does not mark a regular run complete when verification remains unresolved.
• Ran pytest -q tests/test_state_runtime.py tests/test_bounded_execution.py and the tests passed (46 passed).

---

Codex Task