JonKolbert pointed me to the following diff:
A global AbuseFilter that does something like norm(edit_diff) irlike "encyclopediasupreme" should have caught this, but failed to.
The characters being used there are U+1D41E MATHEMATICAL BOLD SMALL E et al. Those were added to equivset.in (which feeds AntiSpoof and AbuseFilter, according to the docs) in rMLEQ4464b4454b48fe2d79b6b84f0810394d6db6b776.
However, this doesn't seem to have ever been deployed, or else something is choking. I visited Special:AbuseFilter/test and tested the following filter against a dummy edit:
ccnorm("𝐞") irlike "e"
Which did not match. I then tested the filter: (U+FF25 FULLWIDTH LATIN CAPITAL LETTER E)
ccnorm("Ｅ") irlike "e"
So I suspect that either the changes in the linked diff above were never deployed, or that something in AntiSpoof or AbuseFilter is failing to handle character equivalencies outside of the basic multilingual plane.