Looks like there is a fundamental tradeoff for startup time here which might prevent using racket directly as a subshell (there are ways around this such as making the racket process a daemon). I'll be interested to see if the 'rough unscientific benchmark' numbers can reach the chez levels and to see some more scientific benchmarks (I've been trying to get the r7 repo to build to run some personal benchmarks, but haven't really had time).

Fantastic work and looking forward to what the future holds!