The issue was that the Python keyframing code was already resolving the
RNA path fully to the owning ID, but then erroneously passing the
non-ID RNA pointer (in this case the sound sequence pointer) to the
keyframing code with that fully resolved path.
Notably, it wasn't just the VSE keyframing that was broken: keying any
non-ID structs via the Python API was broken. A good example is pose
bones: the Python keyframing code was resolving e.g. "location" on
the bone to "pose.bones["Bone"].location", but then passing that
path along with the pose bone struct to the keyframing code. Since
that fully resolved path of course doesn't exist on the bone itself,
keying would fail.
This fixes it by simply passing the owning ID's RNA pointer instead,
which it should have been doing in the first place.
Pull Request: https://projects.blender.org/blender/blender/pulls/123719
Nathan Vegdahl
38f889d07a