Thanks for the well-written report. It looks like we introduced a bug here in our ModelingToolkit v4 release. I fixed it in https://github.com/SciML/ModelingToolkit.jl/pull/677 and will merge and tag the patch tomorrow morning. Also, I made this model into a regression test to catch this in the future.
P.S. The speedup you get from this is before:
0.000285 seconds (4.43 k allocations: 197.484 KiB)
after:
0.000060 seconds (73 allocations: 13.453 KiB)
mostly due to the allocation of the Λ.