rerun and replace dataset in collection fail to coordinate multiple outputs

When running a tool on a collection of inputs, outputs will be a collection. If each job produces multiple outputs, the results will be multiple collections. In the latter scenario, rerun and replace a failed a dataset fail to make other outputs replaced in their collection. I'm not sure my statement is generalized right, since the other output failed me was a collection from `discover_datasets`.

The issue names no file or test. Start by reproducing rerun and replace with a collection whose jobs produce multiple outputs, including an output from discover_datasets; trace how the failed dataset and its sibling outputs are coordinated. Done means all related outputs are replaced consistently in their collections.

Written by the indexing model from the issue text.