fix: require initial tool choice for evidence runs

[1aal]

Summary

• map Runtime run requests with requires_tool_evidence=true to a require_until_first_tool policy
• force tool_choice=required only while the current turn has no tool result, then release back to default so final summaries can complete
• add focused regressions for required tool choice release and Runtime request policy mapping

Validation

• uv run pytest tests/run_agent/test_required_tool_choice_policy.py tests/runtime_manager/test_registry.py -q
• uv run pytest tests/run_agent/test_notice_spine.py -q
• uv run ruff check agent/conversation_loop.py agent/agent_init.py run_agent.py runtime_manager/manager.py runtime_manager/worker_main.py tests/run_agent/test_required_tool_choice_policy.py tests/runtime_manager/test_registry.py
• python3 -m py_compile agent/conversation_loop.py agent/agent_init.py run_agent.py runtime_manager/manager.py runtime_manager/worker_main.py
• git diff --check
• git diff --check origin/apecloud-base

[1aal]

Closing per owner decision: do not proceed with the Runtime required tool_choice approach for evidence runs. We will revisit the product/runtime contract before making native Hermes Agent changes.