Behaviour implemented by coding-agent provider adapters.

Lifecycle and command callbacks run in bounded, session-owned tasks. capabilities/1 runs directly in the SessionServer and must return without blocking. Long-running I/O belongs in provider-owned processes which publish normalized messages through AgentHarness.Provider.Sink.

A PID handle is monitored automatically. For an opaque handle, return a runtime PID as session_info.monitor; otherwise the adapter must report transport loss with AgentHarness.Provider.Sink.transport_down/2. A runtime tied to the logical session should monitor sink.pid, not the temporary process executing open_session/2.

If start_turn/4 cannot determine whether upstream work began, return {:error, {:turn_start_uncertain, reason}}. AgentHarness records a terminal failure and retires that provider session so potentially running work cannot be mistaken for a reusable idle conversation.

If respond/3 or cancel/2 cannot determine whether the command reached the provider, return {:error, {:provider_command_uncertain, reason}}. A plain {:error, reason} from respond/3 means the response was definitely rejected and may be attempted again. Cancellation failures retire the provider session because the upstream turn may still be running.