log/sqrt for Symmetric/Hermitian arguments, and acos/asin/acosh for Symmetric/Hermitian arguments with Complex eltype, are type unstable. Specifically, these functions return Symmetric/Hermitian where the argument is positive semidefinite and an unannotated matrix otherwise.
Similar functions accepting, e.g., StridedMatrixs always return an unannotated matrix, even if the result internally received annotation. The question is whether the type-unstable methods above should perhaps do the same, or whether instead the type-stable methods should join the type-unstable methods in preserving annotations.
The inconsistency was surprising when working with these methods. Otherwise I see arguments for each choice.
Thoughts? Thanks!