Although this looks like a bug in SuiteSparse, it is tricky. The ldlt wrapper is indeed invoking a CHOLMOD function with arguments requesting the indefinite (simplicial) algorithm. However, to preserve sparsity that algorithm sacrifices stability - it does no pivoting. Thus it may encounter zero divisors unless there are a significant number of non-zero entries on the diagonal or a tiny shift value is provided. These latter conditions will produce a factorization without error flags, but it may not be useful, i.e. it may have large forward and/or backwards errors when used in a solver.
In practice many indefinite sparse matrices have strong diagonals, so the algorithm works for them. If your problems provoke the error condition, you should check their character - you may be lucky that black-box code use did not just produce nonsense instead.
SuiteSparse-CHOLMOD advertises itself as a library for positive-definite systems, and the indefinite case is in the category of “this might work but caveat emptor”. So I don’t think it’s a bug, but it is a challenge to document properly.