I ported Picomatch to Rust. It passed 1,977 tests and lost the benchmark by 18x

A Port Mortem 2026 write-up about a JavaScript to Rust port, one strange Unicode character, and what green tests do not tell you.

The first benchmark result was not flattering.

Picomatch, running under Node, handled about 7.18 million matches per second in my test. The direct Rust implementation managed about 395,000. That made the Rust version roughly 18.2 times slower.

I kept the result because it makes the point of the project clearer. The hard part was verification: how much evidence would it take to trust an AI-assisted rewrite beyond the obvious cases?

The small library that was not a small target