Jido.AgentServer.Lifecycle implementation that flushes the active
Gralkor session when an agent process terminates.
Sole job: on terminate/2, read the committed Jido thread id from
state.agent.state[:__thread__].id and fire-and-forget
Gralkor.Client.flush(thread_id) so the buffered turns don't strand
in Gralkor.CaptureBuffer state outliving the agent. Failures are
logged but never block termination. First-turn agents (no thread
committed yet) terminate without calling Gralkor.
Memory consolidation while the agent is alive is owned by
JidoGralkor.ContextRotator, not this module. The AgentServer's
built-in :idle_timeout knob is unaffected — consumers can use it
directly if they want a reaper, but this module no longer wires one
on their behalf.