Send low-signal dictation to speech backend

[jefking]

Summary

• stop blocking near-silent browser recordings before transcription
• still submit captured audio so backend logs include speech stats and Whisper can attempt recognition
• show a clearer quiet-input message only when transcription returns no text

Validation

• go test ./internal/web
• go test ./...
• Playwright silent fake-microphone smoke against local source server returned: No speech was recognized. Check microphone input.