feat: batched tick-level inference (llm:chat-all)

Roadmap priority: B4 — described in the progress report as "the single biggest simulation-speed win for population-scale models."

Summary

Add a batched, tick-level inference primitive — an llm:chat-all-style call that fires every agent's request concurrently with one barrier per tick, instead of N sequential per-agent calls.

Why it matters

The dominant cost in a population-scale LLM model is wall-clock, not tokens. ask turtles [ let r llm:chat ... ] serializes: 100 turtles at 2s per call is over 3 minutes per tick. Firing them together and joining once per tick collapses that to roughly the slowest single call.

llm:chat-async already exists but retrieval is a blocking runresult, so the modeler has to hand-roll the two-pass pattern (fire all, then collect all) and gets no concurrency control.

What needs to be done

1. Decide the API shape. Options: a command that takes an agentset and a prompt reporter, a two-phase fire/collect pair, or a batched reporter returning a list aligned to the agentset.
2. Respect NetLogo's execution model — results must land back on the right agents, and the barrier must complete before the primitive returns.
3. Per-agent history must stay correct under concurrent commit (the historyLock / commitExchange path from 14f7c35 already handles this; verify it holds at batch scale).
4. Per-agent failure isolation — one failed call must not abort the batch.
5. Must compose with throttling (#48): "fire all" needs to mean "queue all, run at most N concurrently."

Open questions

• What is the natural NetLogo idiom here? A primitive taking an agentset and a prompt-producing reporter is powerful but unusual for this codebase.
• Should the result be assigned to a turtle variable automatically, or returned as a list for the modeler to distribute?
• Interaction with llm:chat-async — does this replace it, or sit alongside?

Related

• #48 throttling (hard dependency — unbounded batch fan-out is exactly the failure mode documented there)
• #49 token tracking (needed to report batch cost)

Start by reading the existing llm:chat-async path, especially the blocking runresult behavior, and inspect the historyLock/commitExchange path referenced in the issue. Resolve the API and result-assignment design with #48 throttling and #49 token tracking before implementation; done means a barriered batch preserves per-agent history, isolates failures, and limits concurrency.