This is oversimplifying. Pipeline stalls aren't the same thing as "horrible memory patterns". Not all workloads can be optimized like that at all, some very legitimate workloads are inherently random access (think RAM caching and databases!) and we still want to run them on CPUs tuned for wide issue of in-register instructions.

And when your very legitimate workload does stall, inevitably, it's nice if the CPU has another thread able to issue while it waits. In fact hyperthreading benefits almost all but the most heavily tuned workloads. Pretty much everyone stalls on a routine enough basis to make it worthwhile.

Now, in the post-spectre world where you need to spend transistors on firewalling/tagging to prevent information leaks, maybe it's not as big a benefit. But it's mischaracterizing history to claim that it was never worthwhile.