fix(runtime): convert_gguf.py vocab fallback is dead-coded with if False

[kurosawareiji7007-hub]

Summary

runtime/tools/convert_gguf.py claims to take vocab from token_embd.weight when writing config.txt, but the assignment is disabled:

if "token_embd.weight" in ten:
    cfg["vocab"] = int(ten["token_embd.weight"].shape[-1]) if False else cfg["vocab"]

The C++ GGUF path always prefers the embedding tensor (emb->dims[1] in runtime/examples/qwen3_gguf_config.h / qwen35.cpp). The Python converter therefore can write a wrong vocab= into config.txt when metadata is missing or disagrees with the tensor, which breaks downstream load_weights / LM-head sizing for this tool path.

Expected

Prefer vocab from token_embd.weight shape (ggml dims[1] / Python shape[1] for 2D), matching the runtime.

Scope

No GPU required — tooling correctness only.