Why the edits feel slow, measured through the production relay with the reasoning switch deployed: OpenRouter's messages endpoint ignores every reasoning shape for DeepSeek v4.1 flash (effort none, enabled false, a 64-token cap all still think 2–8K tokens), and the sweep of allowlisted alternatives lost on output quality or speed — Luna drew a static thief in 12.7 s, Qwen invented a tap-to-start in 41 s, v4 pro took 49 s, chat v3.1 58 s. DeepSeek flash keeps the best animated edits, so the thinking stays. What changes: the thinking already streams to us from the first byte, so AcServer now emits item/reasoning/delta for thinking_delta events (never kept in the conversation) and the engine lets it fly by in the pending-row ticker, dimmed and plain, until code arrives. The wait reads as work. SpeechCapture.swift takes the microphone, recognizer, pitch/energy analysis and release timing out of WalkiewareSession, reporting through onEvent, onLevel and onReplayRelease; the session keeps screen state and the bridge. The gesture UI tests pass on the split. Tests/replay-edit.mjs is the per-round timing harness used above; Tests/probe-openrouter.mjs asks OpenRouter directly which reasoning shape this model honors (needs the vault key, so a person runs it). Build 78. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>