The planner is in fact willing to use hash aggregation when work_mem is
not set high enough for everything to fit in memory.  This has been the
case since commit 1f39bce0, which added disk-based hash aggregation.

There are a few remaining cases in which hash aggregation is avoided as
a matter of policy when the planner surmises that spilling will be
necessary.  For example, callers of choose_hashed_setop() still
conservatively avoid hash aggregation when spilling is anticipated.
That doesn't seem like a good enough reason to mention hash aggregation
in this context.

Backpatch: 13-, where disk-based hash aggregation was introduced.