test: large transcript stress/correctness — repeated rolling compaction

[emesal]

context

the roadmap calls for stress testing with large transcripts. correctness under load is the goal — not performance benchmarks.

scope

integration test (gated on env var, uses free/agentic or free/text-generation):

1. build a synthetic transcript with 80–100 messages (mix of user, assistant, tool_call+tool_result pairs)
2. run rolling compaction repeatedly until message count is stable (≤4 non-system messages)
3. after each compaction round, assert invariants:
   • no orphaned tool results (every tool message has a corresponding assistant tool_call)
   • system messages preserved throughout
   • summary grows or stays non-empty
   • message count strictly decreases each round (or hits the ≤4 floor)
   • transcript anchor written after each round
4. after full compaction: verify final context is coherent (loadable, parseable, no corrupt JSON)

notes

• free models may be slow; set generous timeouts
• if LLM returns empty IDs repeatedly (triggering fallback path), that's fine — test the fallback path too
• seed the transcript deterministically so failures are reproducible
• this test will be slow (~minutes); mark clearly in output