"Upgraded nodes enforce the fork by rejecting blocks that are too large (which legacy nodes continue to accept)."
How is that not a hard fork?
Discussion
The definition of a soft fork is a further constraint. A narrowing of the consensus rules. The definition of a hard fork is relaxing a constraint. A broadening of the rules.
So expanding the main block size limit (Satoshi’s 1MB limit) is a hard fork because it relaxes the limit. Unupgraded nodes would NOT accept the new, larger blocks. It’s a hard fork because everyone must upgrade, or be left behind.
By contrast, reducing the block size is a soft fork. Unupgraded nodes will accept the new, smaller blocks just fine. It’s backwards compatible. People can opt to run old software and continue to use Bitcoin.