The Qwen3.8 GGUF reasons internally before answering and the llama.cpp
endpoint ignores enable_thinking — hidden reasoning consumed the whole
400-token persona_reply budget (HTTP 200, content='') -> LLMError -> live
500 'LLM service unavailable'. The 800-token judge JSON had the same
starve and silently degraded to 'pending', so customers who decided to
buy never closed the session.
- persona_reply 400 -> 1200, evaluate_turn 800 -> 1600
- tests/test_llm_budgets.py pins the floors so a regression to the old
values fails loudly
- full suite: 540 passed; independent review PASS (truncation-safe:
partial protocol JSON is rejected, never rendered)
c5553eb9a6