Align live target/recovery control path with evidence-based execution contracts

[mmprotest]

Motivation

• The runner was prematurely pinning primary execution targets from single-file writes and losing important execution contract information (cwd, evidence), which caused silent drift, verifier/wrapper thrash, and inert recovery behavior.
• Hard failures from actual direct runs were not consistently arming live recovery state used by execution control, leaving repair flows and summaries out of sync with real validation signals.
• Shell normalization on Windows cmd missed many embedded Unix compound fragments and detached-launch assumptions that leak unsafe behavior into execution.

Description

• Introduced an evidence-ranked seeding model and replacement rule via _TARGET_EVIDENCE_ORDER and updated _seed_primary_execution_target so a new seed replaces the current only when it has strictly stronger evidence (e.g., write_only < indirect_validation < direct_validation < direct_run).
• Tracked the primary execution contract as a minimal pair (target, cwd) with helpers _target_contract, _primary_execution_contract, _same_target_contract, and preserved the contract through mission sync and state fields (including _primary_execution_target_cwd and _primary_execution_target_evidence).
• Made live recovery activation evidence- and contract-based so activate_live_recovery_on_primary_failure only arms recovery when a hard failure occurs on the current primary contract and updates the same live state the control path reads (sets _recovery_mode, _failing_file, _failing_target_contract_summary, _primary_target_minimally_valid etc.).
• Added narrow recovery-time guards in small_model_tool_guard to block sideways target-switch and new-wrapper thrash during recovery when the primary target is not yet minimally valid, while still allowing reads, helper edits, and reruns of the same contract.
• Refreshed/augmented live validation candidates after material mutations with refresh_live_validation_candidates and called it from tooling flows so the validation set is not stuck on startup-only defaults.
• Tightened Windows cmd shell checks in classify_and_rewrite_command and _detect_invalid_cmd_detached_launch to block embedded tail/wc fragments and more &-based detached-launch assumptions using conservative pattern checks (no full shell parser).
• Fed a compact recovery block into the model context packet via build_model_context_packet so the model can see the pinned primary contract, recovery_mode, a short failing summary, minimal-valid flag, and whether switching is blocked.
• Kept completion honesty by prepending an incomplete/execution-status notice when finalization runs while the primary target lacks clean direct validation.

Testing

• Added/updated unit tests covering evidence replacement, cwd-aware contract behavior, recovery activation gating, recovery drift blocking, embedded cmd patterns, validation-candidate refresh, and an e2e-style regression for scaffold→real-target failure drift.
• Ran pytest -q tests/test_state_runtime.py tests/test_state_tooling_mutation_policy.py tests/test_shell_command_guard.py and observed all tests passing: 83 passed.

---

Codex Task