What is wrong with my block GMRES implementation?

After a good look at "Iterative methods for sparse linear systems" (also by Y. Saad), I found out my problem.

One needs to perform the QR decomposition of R, as V0, R0 = np.linalg.QR(R) (as it was done), use V0 as the first block in the Block Arnoldi procedure (as it was also the case), but the e1 vector needs to be set using R0, to e1[:s, :s] = R0, instead of using the norm of R as incorrectly done in my code above (which would work in certain cases).

For those who are interested, a working implementation is available at: https://gist.github.com/pierre-24/d638dcb908e28cde759c04989be7b603