Measured on an M5Stack Cardputer: 67 tokens 2514 -> 1056 ms, 126 tokens 5687 -> 2281 ms, 12 tokens 528 -> 245 ms; same heap. Bit-identical: 4-token PIE kernel with the tile loop in assembly, attention loop restructuring, lrintf-free rounding, LayerNorm/quantisation on both cores, 64-byte-aligned embedded model. Within float rounding (outputs move <= 0.0035): fc2 in 256-neuron groups, Cephes exp in the attention softmax, heads read each matrix once (u = W^T q). Host conformance builds with -ffp-contract=fast like the device GCC: max|dp| 0.0163, no pick or span flips.
Rust: no Instant::now trap on wasm32-unknown-unknown (discussion #1), 4-token tiled matmul (~18% faster, same output), make_protos length checks, wasm/simd128 notes. All engines: noul/span questions reject options; protos without a center centre on the count-weighted prototype mean. JS/Python/Rust: tokenizer stops after the tokens the model reads. Conformance fixture regenerated; every engine passes.
Adds python/ (numpy), rust/ and esp32/ (ESP-IDF component and Arduino library), conformance/ with reference outputs every engine is tested against, load.js, corrections.js and TypeScript types, and assets/promo.mp4. tinydecide.js now rejects choice/score questions without 2 to 32 options and treats a null lam as 1. Model files are unchanged.