[JNK234]

Closes #53.

What it does

Every agent in a set asks the LLM its own question at the same time, and the model waits once for all the replies. ask turtles [ llm:chat ... ] waits for each turtle in turn; llm:chat-all takes about as long as the slowest single reply.

let results llm:chat-all turtles [ -> (word "You have " energy " energy. Graze or rest?") ]
;=> [[(turtle 0) "Rest." true] [(turtle 1) "Graze." true] ...]

let decisions (llm:chat-all turtles [ -> (word "Size " size ". A predator appears.") ] schema)
;=> [[(turtle 0) [["action" "hide"] ["confidence" 0.8]] true] ...]

• The prompt reporter runs as each agent, like ask, so it reads that agent's own variables.
• Each entry is [agent reply ok?]. A failed call only marks its own entry false, carrying the error message; the rest of the batch still completes.
• The optional schema reuses the structured-output path (llm:chat-with-schema), so each reply arrives as NetLogo values.
• Built on what is already there: each agent's profile, the provider throttle, 429 retry, per-agent history and token usage all apply, with no new configuration.

Design notes

• Prompts first. Every prompt is built before any request goes out, so an error in the prompt reporter costs no API calls.
• Waiting. The batch keeps waiting while replies keep arriving, so later agents queued behind a throttle cap still finish. It gives up only after a full timeout budget with no reply at all, so a hung provider costs one budget, not one per agent.
• History safety. A reply that arrives after the batch gave up on it, or after the wait was halted, never writes history.
• Refactor. Schema validation moved into a shared parseObjectSchema helper; llm:chat-with-schema's error messages are unchanged.

Verification

• sbt test: 268/268, including 14 new tests.txt cases. They cover triples, prompts running as each agent, per-agent history and usage, profiles, failure isolation, schema parsing and bad schemas, non-string prompts, an empty agentset, concurrency, a hung batch costing one budget, progress extending the wait, and rejecting extra schema arguments.
• Live end-to-end harness against Groq (openai/gpt-oss-20b): 48/48, including the new T11 (6 turtles through a throttle cap of 3, plus a schema batch).
• Live timing with 8 turtles: 4.83 s with ask turtles [ llm:chat ] versus 1.22 s with llm:chat-all.
• The throttle is not asserted by a dedicated Scala spec. chat-all calls the same provider method that ThrottleIntegrationSpec covers, and T11 exercises a cap live.

Docs

docs/API-REFERENCE.md: a quick-reference row, a ### llm:chat-all section, and a "Batched Calls" usage pattern.