If you stick to marcel commands and python functions, then all the processing happens in a single Python process (with a small number of exceptions). Beside the process overhead, this approach also avoids serialization/deserialization costs, and even string copying costs. x | y | z just passes Python tuples from one command to another via function call.
Can you expand on your first point? Why are tabular and hierarchical output so important for your work?
As someone who's leaned into Nushell deeply, I can tell you that the tabular and hierarchical data output is the major selling point of Nushell. For me and my team, who work with data and metadata constantly, it's an incredible productivity boost.
The major pains with Nushell are debugging pipelines that are too slow, having to learn the new language constructs, and process overhead for going in and out of Python for other pieces.