Designing an Attribute-Editing Feature on StyleGAN

Product wants a face-editing feature on top of a trained StyleGAN generator: given a generated face, let the user drag a slider that makes the face look progressively older, without changing anything else about the image (pose, identity, background, hairstyle).

1. Explain why this feature is architecturally possible on StyleGAN specifically, tying your answer to the mapping network and per-layer style injection.
2. Describe, at the level of this subject, how you would actually construct the "age" editing direction, and what it means mechanically to apply it.
3. A teammate implements the slider by editing the raw noise vector z instead of w, reasoning "z is what the user's random seed actually is, so editing z is more direct." Evaluate this design choice.