While your question is pretty general, your perspective is quite limited to C language. Some other languages, such as Go, can have dynamic-sized stack allocation. (Actually, the user does not even bother stack or heap is used in Go)
Take Linux, a giant user of C language, as an example. Allocating a large chuck of memory on stack is just not useful, even given unlimited amount of memory. You note the most critical reason yourself in the post:
(assuming that data doesn't need to be returned outside of the current stack frame)
However, large data structures are almost always for others to use. Just consider sys_clone (that the kernel eventually generates a body for the new thread) or sys_mmap (that the kernel manipulates existing virtual memory area structures from elsewhere). Allocating them on the stack seems pointless.
Take Linux, a giant user of C language, as an example. Allocating a large chuck of memory on stack is just not useful, even given unlimited amount of memory. You note the most critical reason yourself in the post: (assuming that data doesn't need to be returned outside of the current stack frame) However, large data structures are almost always for others to use. Just consider sys_clone (that the kernel eventually generates a body for the new thread) or sys_mmap (that the kernel manipulates existing virtual memory area structures from elsewhere). Allocating them on the stack seems pointless.