Inference-stack review: highest-leverage experiments from ender-inference

Read-only review of ender-inference origin/staging found four reusable performance hypotheses: (1) the Python control-panel hop opens a fresh upstream HTTP connection and forces Connection: close for each LLM stream, while Go Railtail pools the next hop; measure raw Adam vs routed TTFT before replacing it with a cancellation-safe pooled async client. (2) TTS has process-local LRU caching and intentionally allows concurrent identical cold misses, so singleflight plus a revision-keyed shared cache could reduce duplicate GPU synthesis. (3) STT starts/cancels a Qwen0.6B turn-classifier request for each streaming ASR text revision and can reprocess cumulative audio at checkpoints; instrument detector calls/cancellations and processed-audio/source-audio amplification before adding debounce/prefix reuse. (4) a historical RTX5090 Qwen3-1.7B native-engine matrix failed every performance comparison versus vLLM (vLLM reached 12.6k tok/s vs native 672 tok/s at 128-token prompt/c60), so revive native serving only with a materially new kernel/scheduler thesis. These are repository-derived hypotheses; current Adam deployment and Linear benchmark artifacts still require live verification.