Docs: nlu-architecture.md and tool-calling-runtime.md describe old FunctionGemma system

[0xharkirat]

Priority: Medium

Two docs in docs/ still describe the pre-EmbeddingGemma architecture:

1. docs/nlu-architecture.md — Proposes all-MiniLM-L6-v2 (23MB) as the embedding model and regex-only parameter extraction. The actual system uses EmbeddingGemma 308M and Qwen3 for slot filling. Status header has been updated to note this is a historical design doc, but the body still reads as a current proposal.

2. docs/tool-calling-runtime.md — Describes FunctionGemma 270M's 2K context budget and two-tier optimization. The current system does not use FunctionGemma at all.

Options

• Option A: Rewrite both docs to reflect the current EmbeddingGemma + Qwen3 pipeline
• Option B: Keep as historical design docs with a clear header banner: "This document describes the original design. The current implementation uses EmbeddingGemma + Qwen3 — see architecture.md"

Files

• docs/nlu-architecture.md
• docs/tool-calling-runtime.md