A critical assessment of effective core potential (ECP)-based singledeterminant (SD) fixed-node diffusion quantum Monte Carlo (FNDMC) accuracy in prototypical noncovalent closed-shell systems involving d-elements is presented. Careful analysis of biases and elimination of possible bias sources leads to two findings of practical importance for SD FNDMC in these systems. First, in some systems (HCu:HCu, HCu:CuH), SD FNDMC reveals large biases of interaction energy differences (significantly exceeding the target 2% relative error) vs a reliable coupled-cluster CCSD(T)/CBS (complete basis set) reference. Second, the leading error of SD FNDMC with ECPs was attributed to a higher nuclear charge Z of d-group (pseudo) atoms, when compared to sp elements, in line with a previously reported finding that aggregate SD FNDMC bias tends to increase in systems with higher electronic densities. Therefore, SD FNDMC should only be used with caution in systems with a large Z.