refactor: clarify orchestrator runtime boundary

[cafitac]

Summary

• split the orchestrator contract into setup-only and future runtime protocols
• remove misleading lifecycle stubs from Claude, Codex, and Hermes setup adapters
• update refactor docs/changelog to state that Hermit MCP remains the canonical runtime boundary

Decision

• Chose Option B from .dev/refactor/hermes-executor-readiness/pr-05-runtime-delivery-boundary.md
• Hermit runtime task lifecycle stays on the existing MCP server path (run_task, reply_task, check_task, cancel_task)
• OrchestratorRuntimeAdapter remains available as a future protocol, while current wrappers implement only OrchestratorSetupAdapter

Test Plan

• .venv/bin/python -m pytest tests/test_orchestrator_contracts.py tests/test_hermes_orchestrator_adapter.py tests/test_claude_orchestrator_adapter.py tests/test_codex_orchestrator_adapter.py tests/test_orchestrator_prompt_mapping.py tests/test_orchestrator_event_mapping.py tests/test_mcp_tool_handlers.py tests/test_mcp_server.py -q
• .venv/bin/python -m pytest tests/ -q

Results

• targeted: 42 passed in 0.73s
• full suite: 1079 passed in 32.47s

Future work

• only implement adapter-owned runtime lifecycle methods if a real orchestrator needs a non-MCP delivery path