Inconsistency of IO files when running colabfold_search and colabfold_batch with templates locally

Expected Behavior

If one is to use the input option of colabfold_batch for a folder of a3m files with --template argument, the argument --pdb-hit-file should accept a folder with m8 files. Prediction should be done for each a3m and m8 pair, based on names for example.
Alternatively, the default behavior of --templates should be to look for .m8 files with the same name as the .a3m in the a3m input folder, since this is the way colabfold_search outputs its results. If such a file is missing then it should trigger an online search.

Current Behavior

The argument --pdb-hit-file of colabfold_batch can only be a file, therefore one has to run the command for each each alignment separately, which renders the batch in colabfold_batch moot.
If one submits a concatenated m8 file with all template hits, the prediction for each a3m alignment will use all templates listed in the file, regardless of the first column names used.

Context

I'm working on separating the alignment step from the modelling on our server so that we don't swamp our GPU nodes.

Your Environment

ColabFold v1.5.5
Ubuntu 22.04.4 LTS
cuda_12.3.r12.3

Start with the colabfold_batch and colabfold_search entry points and trace how --template and --pdb-hit-file are handled for batch inputs. Done means a folder of matching .a3m and .m8 files is paired per name, with online search used when a matching .m8 file is absent.