Benchmark constrained-decoding throughput

Benchmark the throughput and latency of constrained decoding with Universal Byte-Level Encoding, including the overhead of exposing per-symbol alphabet tags to guided or grammar-constrained decoding engines.

Background

Universal Byte-Level Encoding requires constrained decoders to know whether each expanded symbol belongs to the UTF-8 or UTF-16 alphabet before applying byte-level constraints. The paper describes this as an implementation requirement and explains the decoder state needed for streaming.

However, the reported experiments do not measure the performance cost of constrained decoding. Establishing throughput and latency under guided or grammar-constrained generation remains an explicitly identified unresolved evaluation task.