refactor(native-tools): add Relay-managed tool layer

[nvddr]

Summary

• make xr-ai-nat the toolkit-independent native tools layer
• run new typed tools through NeMo Relay and expose model-backed agents as ordinary tools
• add the generic AgentRunner seam; the bundled Agent remains a bounded default loop, while custom or future framework-backed runners stay behind a normal tool
• forward Relay-rewritten request content and controlled lineage headers to xr-ai-models, supply request/response OpenAI codecs, omit empty tools, and accept tool-call-only content: null responses
• retain existing NeMo Agent Toolkit function groups only behind legacy extras

Validation

• uv run --project tests pytest tests/test_native_tools.py tests/test_models_openai_compat.py tests/test_nat_models_llm.py tests/test_functions_types.py tests/test_docs_versions.py -q
• uvx ruff check agent-sdk/xr-ai-nat/xr_ai_nat/agent_runner.py agent-sdk/xr-ai-nat/xr_ai_nat/agents.py agent-sdk/xr-ai-models/xr_ai_models/_openai_compat.py tests/test_native_tools.py tests/test_models_openai_compat.py
• uv run --project tests --with pyright pyright agent-sdk/xr-ai-nat/xr_ai_nat/agent_runner.py agent-sdk/xr-ai-nat/xr_ai_nat/agents.py agent-sdk/xr-ai-models/xr_ai_models/_openai_compat.py tests/test_native_tools.py tests/test_models_openai_compat.py
• isolated base-install check verified no installed nvidia-nat-* distribution
• uv build --wheel agent-sdk/xr-ai-nat

Follow-up PRs add streaming vision as a normal native tool and route voice plus background triggers through registered tools.

[wenxind-nvidia]

Blocker

• agent.py:135 drops request.headers when calling LLMService.chat(). Relay automatically injects Dynamo session-lineage headers and allows intercepts to add headers, but none reach the actual model HTTP request. Extend xr-ai-models with controlled per-call headers and forward them here. Relay's managed-call contract explicitly requires the provider callback to consume the mutated LLMRequest.

Suggestions

• agent.py:147 supplies only response_codec. Consequently, request intercepts always receive annotated=None, preventing provider-neutral middleware from inspecting or rewriting messages and tools. Pass OpenAIChatCodec as codec too, and omit tools when empty because the codec rejects tools: null.
• agent.py:265 rejects content: null. That is valid for tool-call-only OpenAI responses, so a Relay execution intercept returning such a response crashes before the tool loop. Normalize None to "", matching xr-ai-models.

Nit

• AGENTS.md:61 calls out public worker SDKs but omits the new xr_ai_runtime surface.

Assumption: Relay is intended to instrument and modify the real provider request, not merely emit local lifecycle events. The PR documentation supports that interpretation.

Focused verification: 7 tests passed, Ruff passed, and the package built successfully. Full test-project type checking could not run locally because PyNvVideoCodec has no macOS ARM wheel; PR CI is green.

[nvddr]

Addressed in caca425: Relay now receives request and response OpenAI codecs; its rewritten content and controlled headers reach the model service; empty tools are omitted; and tool-call-only null content normalizes to an empty native message. Added focused coverage for the Relay boundary and a custom AgentRunner tool.