feat(desktop): add renderer-owned Realtime Talk transport

[TheAngryPit]

Summary

Adds an opt-in Desktop Realtime Talk transport that keeps Hermes Core as the single chat authority while using a provider Realtime session only for low-latency speech input and playback.

This advances #44 and depends on the narrow Hermes Desktop composer ownership contract in NousResearch/hermes-agent#100666.

Problem

The plugin's existing Desktop flow cannot safely take microphone ownership while the built-in wake listener is active. It also cannot submit finalized speech or observe the canonical assistant response through a supported renderer contract.

Using the relay's normal automatic-response mode would create a second answer authority: the provider could answer the user directly, independently of Hermes chat, and could receive a provider tool catalog that the renderer transport does not need.

Design

The change is split into two reviewable commits inside this PR:

1. Renderer-only session minting

   • adds a route-authorized desktop-relay session mode;
   • mints provider sessions with automatic_response=false;
   • sends an empty provider tool catalog;
   • preserves the existing route authorization and leaves other Talk surfaces unchanged.

2. Desktop controller lifecycle

   • mounts through the existing composer.actions render contribution;
   • acquires/releases the Core voice lease and respects stale-session fencing;
   • sends finalized user speech through the canonical composer submit path;
   • waits for the canonical Hermes assistant snapshot before requesting provider audio;
   • handles barge-in with the active provider response_id and clears local/provider audio buffers;
   • ignores stale transcripts and stale assistant snapshots across restart/navigation;
   • disposes subscriptions and media resources on stop, failure, and unmount;
   • treats a rejected Core submission as an explicit failure instead of waiting forever;
   • exposes a one-click retry after startup or runtime failure.

Core owns microphone/wake/composer coordination. The plugin owns provider session minting, WebRTC, audio playback, interruption, and UI state.

Security and authority boundary

• no provider credentials are stored in the renderer;
• the existing backend route authorization remains in force;
• renderer-only minting disables provider automatic responses;
• renderer-only minting exposes no provider tools;
• the provider voices only the already-authoritative Hermes assistant text;
• no private operational files or local paths are included.

Testing

Focused Desktop behavior:

• tests/test_desktop_plugin.py — 6/6 passed;
• tests/test_desktop_turn_interruptions.py — 7/7 passed;
• includes the rejected-submit regression.

Broader local validation:

• plugin suite with the single known macOS path-literal test deselected — 1333 passed, 16 skipped;
• the excluded pre-existing test rejects the substring private, which is also present in the standard macOS temporary-directory root; it is unrelated to this change;
• Ruff — passed;
• node --check desktop/plugin.js — passed;
• uv build — passed;
• git diff --check — passed.

Manual validation still required

A live microphone/WebRTC canary requires a Hermes Desktop build containing the Core dependency and explicit microphone permission. It should cover:

1. start and stop;
2. one spoken turn returning one canonical Hermes answer;
3. barge-in during playback;
4. navigation during an in-flight turn;
5. provider/session failure followed by one-click restart.

No live service, installed plugin, credentials, or user configuration was modified while preparing this PR.

Related

• Advances feat: channel-neutral voice transport — Talk beyond Discord #44
• Depends on feat(desktop): expose guarded composer voice ownership controller NousResearch/hermes-agent#100666
• Hermes RFC context: NousResearch/hermes-agent#77111