Resolve =replace= when it lands on an empty target

[julien-gadonneix]

ConfDict.ops.REPLACE is only consumed when the target subtree is truthy:

if self and kwargs.pop(ConfDict.ops.REPLACE, False):

So replacing a subtree that is missing, null or empty leaves the marker in the config as data, which later surfaces as =replace=: Extra inputs are not permitted when the config reaches a pydantic model.

ConfDict({"a": {"x": 1}}).update({"a": {"=replace=": True, "y": 2}})  # {'a': {'y': 2}}
ConfDict({}).update({"a": {"=replace=": True, "y": 2}})               # {'a': {'=replace=': True, 'y': 2}}

This regressed in #310; before it, the marker was always popped. It also sits oddly with #311 ("an empty mapping is a value like any other") — an empty target is a value, so replacing it should resolve.

Dropping the self and guard alone breaks carrying a pending =replace= inside a grid value, so instead ops are resolved when a patch is applied (update) and left pending when one is merely materialized (__init__, from_yaml, list items). to_uid still rejects unresolved ops.

pytest exca/test_confdict.py passes, plus two new tests.

Made with Cursor