Structured output from LLMs without regex gymnastics

Looking for practical patterns to get reliable JSON out of open-weight models in production. We currently use regex post-processing on freeform output, which is brittle. We're evaluating guided decoding (Outlines, Guidance) vs function-calling APIs vs strict JSON schema enforcement at the tokenizer level. What's actually working in prod for you? Token budget overhead, latency tradeoffs, and failure modes when the model refuses to conform.