Voice mode: distinguish direct answers and app actions from coding requests

[meganrogge]

Problem

Voice mode sometimes submits a new coding request when the user intended a lightweight spoken answer or an application action. It can also answer conversationally when the user intended to create or control a session.

This is especially disruptive while another coding request is already running.

Desired routing

Answer or act in the voice backend

• Application state: current session, session counts/status, current model, attachments.
• Application actions: create/focus a session, select a model, attach a file.
• Lightweight general questions that do not require workspace context or durable chat history.
• Examples: "What does polymorphism mean?" or "While that runs, what is the difference between TCP and UDP?"

Submit to the coding agent

• Requests requiring workspace context, tools, repository inspection, or project changes.
• Answers the user is likely to need in the session history.
• Examples: debug a test, explain a compiler error, refactor code, search the repository, review a PR, create a migration, or investigate a build.

Acceptance criteria

• Direct voice answers do not create a chat request.
• App-action utterances invoke the matching frontend tool and do not become prompt text.
• "Start a new session" consistently invokes session creation.
• A lightweight question can be answered while a coding request continues in another session.
• Tool failures and ambiguities cause a clarification or honest failure response.
• The backend routing prompt/tool definitions document this boundary explicitly.

Frontend context and structured action results are supplied by #329493.