We currently don't handle the current situation well:

1. Notebook has multiple image outputs and is normalized (with ./fix)
2. Writer runs only some cells of the notebook, regenerating only some image outputs
3. Writer runs ./fix on the same notebook

Since the normalizer detects images that need extracting, it'll delete the existing image folder and extract the new images. This is a problem as the notebook will contain <Image /> tags that reference images that don't exist.

We should generally discourage only running some cells of a notebook, but our normalization script should be able to handle it.

Until we fix this, the workaround is to either:

1. Re-run the notebook from start to finish (preferably using tox)
2. If that's not an option, restore the deleted images using Git