Fixed missing voices for certain accents, and voice conversations cutting out on long answers

• Restored and expanded the available voices for Argentinian Spanish and Australian English accents. You now have dozens to pick from in each accent, and you can preview them before you choose.
• Fixed some edge cases where the agent got cut off while speaking longer answers, especially closing summaries. It now has room to finish, and picks up where it left off if it ever does run long.
• Fixed voice conversations that could start up without any sound, or stop hearing the participant midway. If a conversation cannot connect at all, participants are now told and offered the text version.
• File attachments now work with OpenAI GPT models. PDFs and text files were not getting processed properly, which caused errors in design sessions.
• We now show accurate token usage in design and analysis sessions. Hover the small dial next to the model picker to see how full the conversation is.

Updated underlying dependencies to keep your workspace secure.