From the
--split-cat-pheno documentation:
* 'omit-most' causes the largest category to be omitted (breaking ties in favor of removing the first-seen category), while 'omit-last' always removes the last-seen category. (It is often necessary to omit one category to avoid creating linear dependence between the covariates, which breaks --glm.)
...
Note that --glm does not require you to preprocess categorical covariates with this command; it automatically executes "--split-cat-pheno omit-most covar-01" on them.
(I just added the last sentence today, so your memory isn't playing tricks on you if you don't remember seeing it earlier.)
As for merge, if there's only "1 variant" with 3+ alleles present, it's probably ".", and you should preprocess all the input files with e.g. --set-all-var-ids or --set-missing-var-ids before retrying.