Fix non-bool approval policy results failing open

[mikemikimike]

Summary

Callable needs_approval policies are declared to return bool, but a non-boolean result such as None was coerced with bool() and treated as approval not required. This pull request makes every non-boolean callable result fail closed by requiring approval, and routes MCP callable policies through the same evaluator so they cannot bypass the fix.

Fixes #4845

Test plan

• Added ordinary Runner regression coverage for None and truthy non-boolean results from async approval predicates; both cases interrupt before the tool runs.
• Added MCP end-to-end coverage for sync and async policies returning None; both cases interrupt and record zero MCP calls.
• Focused tests: 5 passed in tests/test_run_step_execution.py, 4 passed in tests/mcp/test_mcp_approval.py, and 3 passed in tests/mcp/test_mcp_util.py.
• make format and make lint passed; targeted Ruff, format, mypy, and Pyright checks passed.
• The full repository checks were attempted. Full mypy reported 46 pre-existing errors in 8 files, and full Pyright reported 27 pre-existing errors in sandbox/optional-dependency files. The full non-serial test run reached 6616 passed and 98 skipped, with 38 failures and 40 collection errors caused by missing optional dependencies and existing Windows/sandbox/runloop environment issues. The serial test entry point passed.

Issue number

Fixes #4845

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

AI disclosure

This pull request was prepared with AI assistance. The implementation, tests, and verification results were reviewed before submission.

[mikemikimike]

Local focused verification is green (ordinary Runner, MCP approval, Ruff, mypy/Pyright on touched modules). The repository's fork-triggered Tests workflows are currently completed with �ction_required; I cannot approve them from the contributor account. The PR body records the full local-suite optional-dependency and Windows limitations.