v1.9.8

• Telnyx-only: one API key. The AssemblyAI-direct provider surface is removed entirely — onboarding now asks for your TELNYX_API_KEY only, and every cloud path (streaming, batch, fallback) runs through Telnyx. Backed by live latency measurements: Telnyx routed ~635ms average vs ~770ms through the direct path.
• Streaming engine selector. BOLO_TELNYX_STREAMING_ENGINE and BOLO_TELNYX_STREAMING_MODEL unlock all 12 engines on Telnyx's websocket (Deepgram, AssemblyAI, Speechmatics, Soniox, xAI, Google, Azure, and more) with per-engine parameter rules. Default stays deepgram/nova-3; this release's tested pick is flux-multi — lowest-latency Deepgram family, Hindi code-switching, and your vocabulary rides streaming as keyterms.
• Bolo learns your corrections. Fix a misheard word within 15 seconds of a dictation and Bolo learns the pair: saved to ~/.bolo/learned_vocabulary.json (capped at 100), applied on every future dictation at zero latency, confirmed with a "Bolo learned: X → Y" notification, visible under the tray menu's Learned Words and counted on the dashboard. Only clean single-word substitutions are learned; deletions and full rewrites are ignored. Learning works without any extra macOS permission: the watcher reads the accessibility caret context on a fixed cadence, so no Input Monitoring grant is needed.
• Copy-transcript popup when there is nowhere to paste. If no editable field is focused (or the paste fails), Bolo skips the blind Cmd+V, copies the transcript, and shows a popup with the transcript instead — edit it in place, click Copy to take the edited version, or Dismiss. Edits made inside the popup are learned exactly like paste edits, and the popup never steals focus from what you were doing.