Erasure is heralded. Therefore, in the absence of other noise channels you can always correct against it up to the percolation threshold (24.9% for 3D cubic lattice, 50% for 2D squared lattice). However, using different methods of redundant encoding or post-selection, you can reach much higher rates. so in general, the more redundant the lattice, the better it preforms against erasure (but the worse it performs against unheralded noise). For example, see https://arxiv.org/abs/2212.04834, and many others.
for depolarizing noise in your gates and measurements, you need a 3D lattice. If your noise is symmetric (not biased for X, Z or Y noise), so a symmetric lattice (namely a lattice with the same primal and dual syndrome graphs) will perform best. In general, the least edges connected to a node (syndrome) in the syndrome graphs, and the least gates performed on a qubit in the physical lattice, the better the code will perform. Although naively these two requirements are opposite, see in https://arxiv.org/abs/1810.09621 that it is not always the case.
Also note that "a code able to correct t unlocated errors is also able to correct 2t located errors" (https://arxiv.org/abs/quant-ph/0601066)