feat: shared / group memory (blackboard) alongside per-agent history

[JNK234]

Roadmap priority: B9

Summary

A shared "blackboard" memory alongside per-agent history, plus optional injection of nearby agents' state into prompts — for social simulations where agents need common ground.

Why it matters

History is strictly per-agent (WeakHashMap[Agent, ArrayBuffer[ChatMessage]]). Anything shared — a public announcement, a common rumour pool, a group decision record — has to be rebuilt by the modeler with llm:set-history on every agent, every tick.

Several existing demos work around this: color-sharing, social-deduction, and topology-tournament all hand-assemble neighbour context into each prompt.

What needs to be done

1. Named shared channels — write once, readable by many agents. Likely the same mechanism as B8 (feat: observer-level / unbound LLM primitive for narrators and judges #57).
2. Primitives to write to and read from a shared channel, and to inject it into a call alongside per-agent history.
3. Decide ordering semantics: does shared content sit before or after per-agent history in the message list?
4. Optional neighbour-state injection — a helper that formats nearby agents' visible state into a prompt fragment. (mesa-llm does this automatically via vision radius; ours would be modeler-controlled.)
5. Must interact sanely with memory management (feat: agent memory management — consolidation, eviction, and retrieval for per-agent history #47) — shared channels grow too.

Open questions

• Should this be one design with B8 (feat: observer-level / unbound LLM primitive for narrators and judges #57)? A named-channel abstraction plausibly covers observer-level calls, judges, and shared memory in one concept.
• Read scoping — can any agent read any channel, or should there be per-channel visibility rules (relevant for deception/asymmetric-information models)?