v6.0.2

Fixed

• A ctx.invoke child run's events reach an Observer under its parent's session (#491). A
  sink bucketing by conversation now sees the child's run.started / usage.reported /
  run.completed under it, instead of losing them to a process-wide session-less bucket shared
  with every other conversation's children. Live sinks only, and attribution only: the child's
  stored rows stay session-less, so it still claims no session, is still played with its own
  history, and still contributes nothing to the conversation's transcript. Read a stored child
  back through run.started.parent_run_id.
• A workflow sharing a conversation's session_id no longer duplicates its input into the
  agent's next prompt (#490). Only an agent turn is a conversational turn: a workflow or tool
  run passed the same session_id joins that conversation's event stream, so an operator
  surface keyed by the session sees it, but its input is an orchestration argument and never
  reaches the model as something the user said. "Enrich, then answer" on one session is now the
  one obvious path, with no derived <session>|step id to invent.
  One limit, for a conversation that already ran a workflow on a shared session_id under an
  earlier release: that release replayed the workflow's input into the engine's session, which
  the transcript no longer counts, so the two disagree from that message on. Such a session
  reports openai_agents.session_diverged on every later turn and skips crash repair, the
  reconciliation that replays a message the log recorded and a killed process never wrote. Turns
  themselves are unaffected, and nothing is written into a session AgentDeck cannot account for.
  Reset one with await deck.session_for("<session-id>").clear_session(), which drops that
  conversation's engine-side memory (the event log is untouched), or move the conversation to a
  new session_id. A session that never shared a workflow run is unaffected.