The TTFA pipeline

Ten stages from speech_final to first audio out. Cold path runs them serially. Speculation collapses five of them to ~50 ms each on hit.
Default cold · no caches warm 2135over the 1500 ms gate
1500 ms gate
Optimized warm · all speculation hits 625passes with margin
1500 ms gate
STT silence Middleware RAG Memory LLM TTFT Tool dispatch LLM continuation TTS formatter Cartesia first audio TX pacer

Where the speculation saves time

RAG retrieval50 → 5 ms (cache hit)
Memory load20 → 5 ms (cache hit)
LLM TTFT900 → 50 ms (the big one · ~95% hit in clean rooms)
Tool dispatch300 → 50 ms (pre-dispatched read-only tool)
LLM continuation250 → 150 ms (faster post-tool turn)
TTS formatter300 → 50 ms (Haiku formatter pre-warmed)