moe_simd_silu_gate_up_q4_sbf16 kernel missing from Metal metallib (Qwen3-Coder MoE on Apple Silicon)

[sandeepraju]

Environment:

• BaseRT: 0.2.3
• Hardware: Apple M4 Pro, 48 GB
• Model: Qwen/Qwen3-Coder-30B-A3B-Instruct (converted via basert pull --target default-q4)

Error (repeated per request):
get_pipeline: no function named 'moe_simd_silu_gate_up_q4_sbf16' in metallib

Steps to reproduce:

1. basert pull Qwen/Qwen3-Coder-30B-A3B-Instruct
2. basert serve Qwen/Qwen3-Coder-30B-A3B-Instruct --port 8081
3. Send any chat/completions request

Notes:
The 0.2.0 changelog mentions "fused SiLU gate/up for 4-bit experts" was added for CUDA.
It appears the Metal metallib doesn't include the equivalent moe_simd_silu_gate_up_q4_sbf16 kernel.
The model converts successfully but fails at inference time.

---

Additional context: The model was converted via basert pull Qwen/Qwen3-Coder-30B-A3B-Instruct (convert-on-pull from raw HuggingFace safetensors, not a pre-built catalog model). Conversion completed without errors and chose the default-q4 profile. The kernel name moe_simd_silu_gate_up_q4_sbf16 suggests it selected a sbf16-scale variant that only exists in the CUDA metallib. It may be worth checking whether the convert-on-pull profile selector is Metal-aware when picking MoE quant variants.