test: rolling compaction integration — real LLM with free model

[emesal]

context

rolling compaction makes two LLM calls. we need integration tests that exercise the full path with a real model to verify the behaviour end-to-end.

scope

integration tests in crates/chibi-core/tests/ (or inline under a #[cfg(test)] block gated on env var):

• runtime skip pattern: if RATATOSKR_API_KEY (or equivalent key env var) is not set, print a message and return early — do not fail the test
• use model preset free/text-generation via ratatoskr
• build a fake context with 10–15 messages (mix of user, assistant, tool_call, tool_result)
• run rolling_compact
• verify:
  • message count decreased
  • summary is non-empty
  • system messages were not removed
  • tool exchange atomicity held (no orphaned tool results)
  • transcript anchor was written (finalize_compaction side effect)

notes

• free models may return unexpected JSON; test should be lenient about exact message count but strict about invariants (no orphaned tool results, system messages intact)
• document the skip pattern clearly so future tests can follow the same convention