Eval performance regression

I noticed some eval benchmarks that suffered performance regressions on ~~May 23rd~~ March 23rd:

This is quite surprising because compilation time in general (greatly) improved on that day, which we attribute to the change from OCaml 4 to 5. Eval must be doing something specific that OCaml 5 doesn't like, so some good old-fashioned profiling is in order here. (sad Windows noises...)

The issue names no files or tests; start with the eval benchmarks that show the regression and reproduce the comparison between OCaml 4 and OCaml 5. Profile the affected eval work, including on Windows if possible, and consider the investigation complete when the OCaml 5-specific cause and a concrete path to address it are identified.

Written by the indexing model from the issue text.