Fix calibration multi-population input contract

[uriahf]

Closes #290.

What changed

• fixes _make_deciles_dat_binary() so matching probs / reals dictionary keys are paired population-by-population
• allows different populations to have different sample sizes while still requiring lengths to match within each population
• preserves the existing single-population and one-concatenated-model behavior
• adds focused regression coverage for unequal-size populations in regular and time-dependent calibration
• refines the README with an explicit Train/Test calibration example where the populations have different sizes

Root cause

The binned calibration helper treated every probability vector as though it had to span the sum of all population sizes. That conflicted with the natural keyed representation already used elsewhere in calibration:

probs["Train"] <-> reals["Train"]
probs["Test"]  <-> reals["Test"]

Scope

This PR intentionally makes only the changes needed for #290. It does not introduce a new calibration module or normalization layer, and it does not address the separate time-heuristic or integer-horizon issues.

Validation

The regression tests cover unequal-size populations for both ordinary calibration rendering modes and for create_calibration_curve_times(). GitHub Actions runs the repository test and documentation workflows.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-11 15:39 UTC