Hit merging implementation
If I understand correctly, the current implementation converts individual G4 hits to summed ADC values in the digitization (including noise), the cell hits are reconstructed and assigned energies and then merged in the reconstruction. I think the order of merging/digitization should be reversed, so that energies of different G4 hits in the same cell are summed before digitization. In going from the G4 deposited energy (which already has fluctuations in the energy loss and sampling) to cell energies, the primary source of noise is electronic noise which contributes per cell not per hit.
In the current implementation if you had N hits all within the same cell, the current digitization/hit reconstruction/hit merging sequence would give each hit its own independent contribution to the electronic noise, thus the merged hit would have noise that goes like sqrt(N)sigma instead of just sigma.
The solution would be to disable the CalorimeterHitsMerger algorithm and replace it with an analogous one in the digitation that runs upstream of CalorimeterHitDigi that takes the hit collection and produces a new, merged one.