@pmp-p referenced https://bugs.python.org/issue42266 on the IRC channel. For context on that bug, what they're saying is that the Python 3.8 behavior is correct (i.e. prints 2) and Python 3.10's caching breaks it (and it prints 1, because f(o) is caching the lookup of o.x, without using the descriptor).

MicroPython also has the same bug, but also it looks like our implementation of descriptor lookup is also incorrect in this case and it still does the wrong thing with bytecode lookup caching disabled. A simpler repro for MicroPython is:

class Descriptor:
    pass

class C:
    x = Descriptor()

    def __init__(self):
        self.x = 1

o = C()

Descriptor.__get__ = lambda *args: 2
Descriptor.__set__ = lambda *args: None

print(o.x)

In CPython (3.8), the behavior is that the o.x on the last line (as per https://docs.python.org/3/howto/descriptor.html#invocation-from-an-instance) will first see that type(o).x is a descriptor, and use that instead of o.x. MicroPython doesn't do that -- mp_obj_instance_load_attr will return immediately with the result of mp_map_lookup(&self->members...) and only after that look up the type's members.

Additionally, setting the __get__ property on Descriptor will not set MP_TYPE_FLAG_HAS_SPECIAL_ACCESSORS on C, so the extra descriptor behavior will not apply on type(o).

And finally, just like CPython 3.10, if bytecode lookup caching is enabled, o.x will bypass all of this and use the cached value anyway.