bug: failed reporter calls leave orphaned prompts in history

[JNK234]

Problem

Several reporters append the user prompt to per-agent history before the provider call succeeds, but they do not roll that mutation back on failure.

Current behavior

This pattern appears in multiple places:

• llm:chat
• llm:chat-async
• llm:choose
• llm:chat-with-template (already problematic for separate reasons)

In each case, the user-side prompt is added to the mutable history buffer before the provider response is known. If the request fails because of timeout, invalid credentials, network error, or parsing error, the reporter throws an exception but the partial prompt remains in history.

Why this matters

This makes failures stateful in a way the caller cannot easily see:

• a retry runs with extra prompt context that came from the failed attempt
• llm:choose can pollute history with the internal constrained-choice prompt text
• one transient provider error can permanently alter later conversation behavior for that agent

This is especially confusing because the call appears to have failed "cleanly" from NetLogo’s perspective, but the extension has already mutated state.

Affected code

• src/main/LLMExtension.scala (ChatReporter.report)
• src/main/LLMExtension.scala (ChatAsyncReporter.report)
• src/main/LLMExtension.scala (ChooseReporter.report)
• src/main/LLMExtension.scala (ChatWithTemplateReporter.report)

Suggested fix directions

1. Treat provider calls as transactional with respect to history.
2. Build request context from a snapshot and only commit user/assistant messages after success.
3. Add failure-path tests proving that history is unchanged after provider exceptions or timeouts.
4. Consider whether internal prompts used by llm:choose should be hidden from persistent user-visible history entirely.

[JNK234]

Rechecked against ff094c6 on main. This issue is still real, but the scope has changed. chat-with-template no longer mutates permanent history before success, so that path is fixed. The remaining affected reporters are llm:chat, llm:choose, and the newer llm:chat-with-thinking, which all still append the user prompt before the provider call completes and do not roll back on failure.

[JNK234]

Rechecked after pulling 3d1283f on main. The scope is now much smaller. llm:chat and llm:choose now build request context from snapshots and only commit permanent history on success, so those paths are fixed. The remaining affected reporter is llm:chat-with-thinking, which still appends the user message before the provider call and does not roll it back on failure.