I am looking for details on the principle and recommended workflow using the catchSalmon and catchKallisto functions for differential tx-level analysis with edgeR. The manual mentions but does not offer any details on the functions. help() mentions that a per-transcript overdispersion value is estimated which is then used to divide the transcript counts by. Is that all one has to do, followed by the standard (e.g. QLF) workflow? Is there any benchmarking or performance data available on how the method compares to other approaches such as swish or ballgown or sleuth?

Yes, that's all that is required.

No, the method is not published.It does well in our own comparisons but we haven't had time to write anything up. I have not so far had a need for transcript-level DE in my own medical research and I give priority to things that I need for my own work. When we do write it up, the help page will list the references.

In the meantime, the help page is just a placeholder. I won't be pushing the method until we do publish performance results.