There is no need to drop the tsibble class when combining sample paths to produce an ensembled fable. Reducing the row structure over .model can be done directly with summarise().

I would also have placed .model = "ENSEMBLE" in the summarise() call, but there is a bug with tsibble: tidyverts/tsibble#214. Generally speaking, the .model column should not be necessary in a fable (tidyverts/fabletools#96), but many functions in fabletools still assume it exists (like accuracy()) and need updating. The .mean column is also entirely optional, as all required statistics are computed directly from the distribution now.