FlashInfer issue #4035: RFC on JIT/AOT package KPIs - the three-path artifact model and what 'cold' costs

RFC, opened 2026, still open. Useful not for a number but for the STRUCTURE of the build cost, which every engine that JIT-compiles shares. FlashInfer ships three runtime artifact paths: flashinfer-python (core, runtime JIT source generation/compilation, on-demand cubin downloads), flashinfer-jit-cache (optional, prebuilt JIT shared libraries for covered specializations), flashinfer-cubin (optional, pre-downloaded cubin/header artifacts). The tradeoff is explicitly 'package/image size, workload coverage, and startup/warmup overhead'. The RFC's own KPI table names 'Cold first-call or warmup latency per representative API/module, broken down into runtime compile time vs. cubin download time' as a thing that is NOT currently measured - which is itself the finding: a widely used engine has an RFC arguing it should be measuring its cold-compile cost and has not yet published the result.