Both can`t reach 200B, they not designed for this. avif only reached 850B on lena256.png, from ~1-2kb+ avif can lead on psnr depending on vector, but for ssimulacra2 avif only wins 2-22kb range, further quantization steps become too rough avif_q91.avif->33kb, avif_q92.avif->37kb, avif_q97.avif->59kb, avif_q98.avif->71kb. webp can`t compete with avif here.
Beats JPEG/JP2K/JXL/SQZ in PSNR
It uses CDF 9/7 wavelet transform without K-scaling + your choice of arithmetic (EBCOT-lite + BAC) or Huffman entropy coding.
You absolutely correct, it's indeed inspired by Keyj`s minimp3. Float-point used because it's hard do with 4-byte fixed-point type and achieve ISO conformance. We must to use different dynamic ranges for different parts of decoder (i.e. emulate own floating point) or more than 4-byte fixed-point type. Only float-point support is needed, SSE/NEON intrinsics is not required and can be fully disabled by MINIMP3_NO_SIMD.
Encoders usually harder, because, for example, you can't verify it using reference vectors (there no exact reference to compare with). Also encoders like h264 contains big part of decoder as well, because it must reconstruct encoded frame internally for motion compensation.
There already defines to include\exclude features: H264E_MAX_THREADS=n, H264E_SVC_API=1, MINIH264_ONLY_SIMD.
Complete documentation still in progress and subject to change. For now only comments in minih264e.h and minih264e_test.c can tell how to use it.