test: full/manual compaction integration — real LLM with free model

[emesal]

context

compact_context_with_llm_manual (manual -c flag compaction) makes three LLM calls and produces a 3-message bootstrap context. it has zero test coverage.

scope

integration tests gated on env var (same skip pattern as rolling compaction issue):

• use model preset free/text-generation
• build a fake context with enough messages to pass the ≤2 guard
• run compact_context_with_llm_manual
• verify resulting context structure:
  • system message present (if system prompt was set)
  • user message contains the continuation prompt + summary block
  • assistant acknowledgment message present
  • context.summary is non-empty
  • original messages are gone
  • transcript anchor written

edge cases (unit, no LLM):

• empty summary response → returns error
• context with ≤2 messages → returns Ok(()) immediately, no changes
• missing compaction prompt → falls back to default, emits CompactionNoPrompt event
• missing continuation prompt → falls back to hardcoded default

notes

• edge cases can be pure unit tests without LLM
• the LLM integration test just needs to verify the structure is correct, not the content