Thank you for raising this!

By creating isolates in networkx for empty rows/cols, we'd be making an assumption that such nodes exist. (edit: for example, in networkx, creating a graph by adding edges [(0,1), (3,0)] would not create nodes 2 and 4, though I understand this is not exactly apples to apples comparision here).

As such, this could be an additional convenience feature (as an additional kwarg such as create_isolates=True). Definitely something that can be documented.

for example, in networkx, creating a graph by adding edges [(0,1), (3,0)] would not create nodes 2 and 4, though I understand this is not exactly apples to apples comparision here

Indeed these are not comparable concepts. In your example, 0, 1 and 3 are not vertex indices. They are just "names". They might as well be "foo", (2,3) and 3.14.

In comparison, in python-graphblas we have indices: 1st row (index 0), second row (index 1), etc.

The current behaviour is not treating the python-graphblas matrix as an adjacency matrix when converting to a graph. An n-vertex graph has n-by-n adjacency matrix. Instead, a rather idiosyncratic conversion method is used.

We discussed this and #476 in a recent meeting. I don't recall whether we got consensus, but there is an argument to be made that the nodes should always be added or at least be added by default. Thanks for raising this issue @szhorvat!

Converting GraphBLAS objects to and from graph objects from other libraries is something we'd like to tackle this year, so I think this issue is related to #398. Defining node attributes with a vector could also make sense.