Airframe v0.2.9

Fixes:

• batch_count: 0 -> batch_count: 1 in server_inference.rs (was killing QKV threads)
• Struct layout: per-field quant types aligned between Rust/WGSL
• Grammar control module for schoolmarm+grammar integration
• Shimmy API: tokenizer_arc(), eos_token(), im_end_token(),
  fse_control_from_patterns(), trace_callback()
• PPT invariant cage (B1-B3) for regression detection
• Prefer discrete GPU over integrated in adapter selection
  (multi-GPU laptops: picks NVIDIA/AMD over Intel iGPU)
• 357 library tests pass