refactor(simple-vlm): separate agentic and voice vision paths

[nvddr]

Summary

• add a finite Relay-managed look_at_current_frame tool for normal agentic planning
• keep direct voice token streaming in a separate LiveVisionResponder that shares frame acquisition and the VLM service boundary
• forward controlled Relay request headers and redact inline camera frames from telemetry
• remove the unsupported streaming-tool lifecycle and legacy NeMo Agent Toolkit extras

Validation

• focused tests: 59 passed
• full CPU suite: 833 passed, 1 skipped, 12 deselected
• Ruff passed
• targeted Pyright passed with 0 errors

[wenxind-nvidia]

Blockers

1. agent-sdk/xr-ai-nat/xr_ai_nat/live_vision.py:139-143 drops request.headers when invoking VLMService.stream(). Relay request intercepts can inject authentication, routing, and tracing headers, but none reach the HTTP request. Please extend the VLM protocol/client with controlled per-call headers and add coverage equivalent to the existing agent header-forwarding test.

2. agent-sdk/xr-ai-nat/xr_ai_nat/streaming.py:58-64 manually opens a Relay span and then calls the handler directly. nemo_relay.tools.call() does not run tool execution intercepts, so retry, short-circuit, timing, and result-shaping middleware is bypassed. I reproduced this with a registered execution intercept: StreamingTool never invoked it, while tools.execute() did. This conflicts with the documented architecture that Relay owns tool middleware and guardrails.

Suggestion

• agent-sdk/xr-ai-nat/xr_ai_nat/live_vision.py:127 places the complete base64 camera frame in the Relay request. Relay's start event records it in both raw and annotated request data, exposing potentially sensitive frames to subscribers/exporters and substantially increasing telemetry volume. Please sanitize or replace the image payload in emitted telemetry.

The migration itself is cleanly scoped and preserves participant routing and streaming behavior. The 21 focused tests pass locally, and the current GitHub checks are green.