[Juliusolsson05]

Summary

• restore Deepgram streaming as a best-effort live-preview side channel while keeping batch upload as the final transcript source of truth
• forward streaming transcript events from main to renderer again
• show composer dictation bars from audio levels during recording, independent of transcript preview availability

Debug evidence

• Latest relevant dictation log had healthy audio: 275/286 nonzero AUDIO_LEVEL samples, max peak 1.0, average peak 0.807
• Chunks were intact: 187 renderer chunks produced and 187 main chunks received
• Transcript preview was absent: only one TRANSCRIPT event existed, the final committed text at release

Verification

• npm run build
• npx tsx --tsconfig tsconfig.node.json scripts/test-dictation-imports.ts

Note: project-wide tsc --noEmit still reports existing unrelated repo-wide type errors; build and dictation smoke test pass.