Major AI Pipeline Simplification and Refactoring

Routing infrastructure, including routing_decisions table and related services, has been entirely removed as it was unused in production.

The RESPOND mode has been eliminated across the codebase, replaced entirely by the UNIFIED mode for all interactions and logic.

The tool_worker has been deleted as the ACT loop now runs inline within digest_worker, making the separate worker obsolete.

The system introduced a unified_generate() function, replacing the previous ONNX mode-gate and dispatch mechanism for a single LLM path.

Tool discovery was overhauled: the find_tools skill now allows dynamic discovery, replacing the need to pre-inject available tools into the ACT prompt.

The card rendering infrastructure has been removed, shifting tool outputs to plain text/HTML delivered via the output_service.

Telemetry was consolidated into , and several dead context slots and prompt bloat have been trimmed for efficiency.

Voice handling was significantly upgraded, replacing faster-whisper and KittenTTS with Moonshine Voice and Kokoro 82M, unifying runtime to ONNX-only.

Skill architecture was redesigned around the find_skills primitive, and the introspect rewrite now provides structured reasoning state instead of raw telemetry.

The block protocol was implemented to standardize UI output, replacing HTML injection with typed block arrays sent over WS messages.

• Replaced RESPOND mode with UNIFIED across the entire codebase.

• Unified LLM path via unified_generate(), replacing the ONNX mode-gate.

• Removed card rendering infrastructure in favor of the find_tools skill.

• Switched STT/TTS to Moonshine Voice and Kokoro 82M, enforcing ONNX-only runtime.

• Consolidated telemetry into the context slot.

• Redesigned skill architecture with find_skills, updating ACT prompt and introspect.

• Implemented universal JSON block protocol for UI rendering.