Multi-length values

It can be semantically appropriate for a value to span multiple values of a vector. For example, a multivariate distribution with each variate exposed as a value (or analogously, a global model with each 'series' exposed as values). Another (less motivating) example is a time granule alongside another (e.g. for mixed-granularity temporal joins and MIDAS regression).

Indexing would invoke similar rules to the replicates, where the joint value is not dropped until all references are.

Is this in some sense, a hierarchical vecvec? Where a value has sub-indices that reference positions/attributes within each value.

Start by reviewing the existing vector-of-vectors representation and the indexing rules for replicates. Determine how a value could span multiple underlying values or expose sub-indices, then define how joint values are retained until all references are dropped. Done means the representation and indexing behavior are specified clearly enough to implement and test.