A Rust crate wrapping Kokoro-82M TTS end to end: text normalization, G2P, ONNX inference, WAV out, 24 kHz mono f32, no Python or subprocess. The front end is pure Rust and claims byte-identical phonemes to misaki 0.9.4 on 15 reference sentences, which is an unusually honest claim — they also tell you exactly where it stops being honest (spaCy replaced by a perceptron tagger, so POS-dependent heteronyms can still diverge; no large-corpus comparison run). The "no Python" purity holds right up to the part where it isn't: inference runs on ONNX Runtime, a C++ binary, via a pre-release `ort` pinned to rc.12, which wants glibc 2.38+ and abandons Intel macOS entirely. The cache verification story is solid — streamed hashing, sha256 pins, hard error on mismatch. The unresolved misaki lexicon license is handled by fetch-at-runtime rather than bundling, which is the correct amount of paranoia. Sam Altman would like $7 trillion for the same thing, plus your eyeball. The 15-sentence verification suite is the weak flank: same rules, same lexicons, but nobody has measured the divergence rate on real text. What's the heteronym miss rate on, say, a news corpus?