feat: observer-level / unbound LLM primitive for narrators and judges

Roadmap priority: B8

Summary

An LLM call that is not bound to a turtle's conversation history — for narrators, judges, referees, and environment generators.

Why it matters

Every chat primitive binds to context.getAgent and commits into that agent's history (LLMExtension.scala:57, commitExchange). When called from the observer, the observer becomes the history owner, which works but conflates two different uses: an agent having a conversation, versus the model asking the LLM a one-off question.

Concrete cases from existing demos: a judge scoring a round in social-deduction, a narrator summarising what happened this tick, generating an environment or scenario at setup.

Was raised in discussion with Can Gurkan (Apr 2026) as an observer-level primitive.

What needs to be done

1. Decide the semantics: stateless one-shot (no history at all), or a named history channel the modeler controls.
2. Named channels are the more general design — llm:chat-in "judge" "..." with its own history, independent of any agent. Would also cover shared/group memory (B9).
3. Primitives for clearing and inspecting a named channel, mirroring llm:history / llm:clear-history.
4. Ensure it works when called from within ask (i.e. does not silently pick up the calling turtle's history).

Open questions

• Is a named-channel API a superset of both this and B9 (shared memory)? If so these should probably be one design.
• Does the observer's current history behaviour need to stay for backward compatibility?

Start by reading LLMExtension.scala:57 and the existing llm:history and llm:clear-history primitives, then inspect how observer calls and ask invoke the current context. Compare the proposed one-shot and named-channel semantics, including backward compatibility for observer history. Done means the API design is decided and its behavior inside ask is specified.