Currently, https://immersive-web.github.io/webxr/#xrreferencespaceevent says:

The transform attribute describes the transform the referenceSpace underwent during this event, if applicable.

See also the corresponding explainer section https://github.com/immersive-web/webxr/blob/master/spatial-tracking-explainer.md#reference-space-reset-event

I think this is a bit ambiguous and could be interpreted in incompatible directions, can this be rephrased more precisely? And which direction should the transform go?

How about something like this for the explainer?

The reset event's transform describes the position and orientation of the post-reset reference space's origin relative to the pre-reset reference space origin. For example, if the reset event describes a correction to the floor height in a local-floor space, where the new floor level is 1m above the old floor level, the reset event's transform would have a position of (0, 1, 0). For a tracked object, the pose transform's Y position will be 1m less after the reset since it's now reported as being closer to the (adjusted) floor.

And this corresponding spec text?

The transform attribute describes the position and orientation of the post-reset reference space relative to the pre-reset reference space.

(I don't have a strong opinion about the direction of the transform. This seemed most natural and close to how originOffset works.)