Aggregate reliability metrics over a cohort of versions (N-way).

Where 'dow compare' contrasts exactly two versions, 'dow aggregate' runs your own
N-way aggregators over a whole cohort - K seeds, judges, prompt wordings, or
clustering permutations - selected by an explicit list of versions, a --tag, or
(by default) the spec's entire history. Each aggregator receives every member
aligned per item through its captured payload, which is what a reliability
coefficient over K raters needs (ICC, Fleiss/Gwet, Krippendorff's alpha, ...). dow
ships none of the coefficients; it only wires in the callables you list under
evaluation.aggregators.

With --plot, the spec's evaluation.plots functions are run over the aggregated
results and dow stores each figure they produce as a content-addressed artifact.
dow ships no plotting library - you plug in your own plot functions.

Each run is saved as a durable, git-tracked result bundle (aggregator values +
figure references + provenance); list them with --list and print one with
--show <id>, so a robustness check becomes a citable, reproducible object.

@examples
dow aggregate
dow aggregate v1 v2 v3 --plot
dow aggregate --tag seed
dow aggregate --list
dow aggregate --show a1
