vinicius-piassa/rinha-backend-2026-asm — reverse-engineered prompt

Build me a super fast Rinha de Backend 2026 submission that runs as a tiny fraud detection HTTP service. I want it to listen on port 9999, accept JSON payment requests, compare each request against a big reference dataset, and return a simple fraud score response as fast as possible.

Make it intentionally extreme and low level, written in pure NASM assembly for Linux x86 64, with no framework or libc. Use one small load balancer process and two worker processes in Docker Compose, staying inside the 1 CPU and 350 MB contest limits. Include an offline index builder that reads the compressed reference data and creates binary index files so startup and request handling are fast.

Focus on performance over readability, using Linux syscalls, io_uring, SIMD where useful, prebuilt HTTP responses, and careful parsing. Add build scripts, Dockerfile, compose file, and a profiling script so I can build the image, run the stack, and test it with the contest benchmark.