fix(agent): verify exhaustive result requests

[ericleepi314]

Summary

• add a completion-audit contract to the core task prompt
• require one bounded tool-backed audit when a user explicitly requests all/multiple/exhaustive results
• preserve the audit latch across tool rounds and add regression coverage

Root cause

On terminal-bench 2.1 chess-best-move, clawcodex found e2e4 and stopped despite the request to return all winning moves. Claude Code mechanically enumerated candidates and also found g2g4. A prose-only follow-up audit repeated the same error, so the runtime guard now requires tool-backed validation for mechanically enumerable state spaces.

Verification

• pytest -q tests/test_query_loop_wires_round3.py tests/parity/test_query_state_parity.py tests/parity/test_error_recovery_flow.py tests/test_system_prompt_full.py tests/test_query_extended_thinking.py tests/test_effort.py — 125 passed
• terminal-bench 2.1 terminal-bench/chess-best-move, Claude Opus 4.8, effort high — reward improved from 0.0 to 1.0
• successful local job: eval/harbor/jobs/chess-rerun-tool-backed-audit