My own take is that mixing digest algorithms is permitted, with each descriptor having a single digest. A manifest with multiple descriptors is not required to be consistent between separate descriptors.

That feels useful to me for a few reasons:

• An index may be assembled from different sources, and upgrading systems to use a different algorithm may not be an atomic operation.
• An index could feasibly be generated with fallback entries to support older runtimes with the older algorithms, while allowing new runtimes to pick a more efficient or secure algorithm.
• Layers in an image manifest may be inherited from a base image, and there is value to reusing the same digest + algorithm to support blob deduplication and cross repository blob mounts.
• Artifacts may assemble content that is not homogeneous, and different algorithms may be preferred for different types of content.
• The empty json descriptor has a predefined digest and algorithm that tooling may prefer to leave unchanged even when other parts are migrated.

Realistically, the next digest algorithm of interest is blake3. So the question is do we and how do we mix and match sha256 and blake3 blobs?