I repeat what I wrote previously:
If you do not use the process "Uniform list of channels" anywhere, then the rows of EEG data (from the F matrix of the subject-level ERPs) that you want to average together are not aligned. In you grand average, the original subject-level ERPs would be averaged as if you had not changed the channel file.
If you don't call this process, you have to edit manually the subject-level ERPs to permute the channels left-right in theFmatrices of the flipped subjects.
The function bst_avg_files.m does match the signals by name, if averages the "F" matrices as is.
If you only modify the channel files, without rewriting the data files (i.e. permute the rows of the "F" matrix), then your grand averaged cannot have the the appropriate subjects flipped left-right. If would produce the same average as if you had not flipped anything.
But maybe your step "2. Analysis pipeline" includes a call to "Uniform list of channels"? Or a manual modification of the F matrices?
Anyway, if you get the results you need, this is what matters.