Align live validation refresh, recovery context, and primary target contract seeding

[mmprotest]

Motivation

• Prevent validation state from remaining stuck at startup fallbacks after the repo or live runs produce stronger evidence.
• Make the model explicitly aware when the runner is in recovery on a pinned primary execution contract so it repairs rather than drifts.
• Ensure primary execution target tracking is a minimal (target, cwd) contract that is seeded by evidence and replaceable only by strictly stronger evidence.

Description

• Add augment_validation_config_with_live_commands(...) in villani_code/project_memory.py to persist compact live-observed validation/run commands into .villani/validation.json.
• Extend villani_code/state_runtime.py with _suggest_live_validation_commands(...), accept observed_commands in refresh_live_validation_candidates(...), and call the refresh/augment path from both run_verification(...) (successful verification artifacts) and the Bash direct-run branch so observed direct runs seed live validation augmentation.
• Tighten villani_code/state_tooling.py so write/patch handling only seeds _current_verification_targets with validation-capable candidate file types and call refresh_live_validation_candidates(...) with those candidates instead of every touched file.
• Keep and use the (target, cwd) primary execution contract and evidence ordering (write_only < indirect_validation < direct_validation < direct_run) in _seed_primary_execution_target(...) while allowing strictly stronger evidence to replace earlier seeds.
• Inject a compact recovery block into the model context packet in villani_code/context_projection.py only when recovery is active, including primary_execution_target, primary_execution_cwd, primary_execution_evidence, recovery_mode, failing_target_summary, primary_target_minimally_valid, and switch_blocked, and render it as a compact single-line contract for model prompts.
• Add and update focused tests to exercise: live validation augmentation from observed commands, recovery block suppression/injection and inclusion of both target and cwd, and an integration flow for scaffold write → direct run → stronger evidence replacement → recovery context → validation augmentation.

Testing

• Ran targeted tests: pytest -q tests/test_state_runtime.py::test_model_context_packet_omits_recovery_block_when_not_recovering tests/test_state_runtime.py::test_model_context_packet_includes_compact_live_recovery_contract tests/test_state_runtime.py::test_live_validation_candidates_refresh_after_material_changes tests/test_state_runtime.py::test_single_file_write_seed_is_replaceable_by_stronger_direct_signal tests/test_state_runtime.py::test_primary_target_contract_distinguishes_same_file_different_cwd tests/test_state_runtime.py::test_hard_failure_activation_requires_same_primary_contract tests/test_state_tooling_mutation_policy.py::test_integration_scaffold_then_direct_run_reseeds_primary_and_live_validation and all passed (7 passed).
• Ran broader suites including tests/test_state_runtime.py, tests/test_state_tooling_mutation_policy.py, and tests/test_mission_state_runtime.py and observed 83 passed for the invoked scope; all automated tests in the exercised suites passed.
• The changes are covered by unit and an integration-style regression test added/updated in tests/test_state_runtime.py and tests/test_state_tooling_mutation_policy.py, and no unrelated pre-existing failures were introduced.

---

Codex Task