I thank and agree with my fellow responders @steven and @dave.allured.
Maybe the path of least resistance might be a “dimensions” attribute (NetCDF 
) whose type is a variable-length sequence of the string-pair compound (which presumably holds the quantity name or symbol and the unit designation). The length and order of elements of the sequence would be expected to be in sync with the leading dimension of your extendible dataset. The only quirk would be that appending a new dimension would mean that you must H5Aread the whole sequence, extend it (in memory), and then H5Dwrite the whole attribute again. Not pretty, but, unless your are constantly changing dimensions, not really a performance bottleneck for a few string-pair records.
Best, G.