Release v1.1.0 · moldyn/MPP

New Features

LumpingKernel seed parameter (TASK-7.1): LumpingKernel now accepts a seed argument (integer, passed to numpy.random.default_rng) for reproducible stochastic runs. In YAML configs, set stochastic.seed: <int>.
New quality metrics (TASK-5.5/5.6/4.5.1-4.5.3): Two new properties on Lumping:
- Lumping.silhouette — sklearn silhouette score
- Lumping.calinski_harabasz — sklearn Calinski-Harabasz index
Also adds a --metrics CLI flag that prints all quality metrics as key=value pairs.
source config key is now optional: The source field in YAML configs no longer needs to be set. When present, it is resolved relative to the config file.

Stable sort in LumpingKernel: Switched to kind='stable' argsort to eliminate non-determinism when states have equal similarity scores. This affected Python 3.13/3.14.
Sankey diagram scaling: Gap and label threshold are now relative to dataset size.
Transition matrix heatmap: Removed spurious % suffix from annotations.
Plot font: All plots now use Latin Modern Roman (via text.usetex=True) for consistent typography.

Change	Impact
Stable-sort algorithm — For systems with tied similarity scores, the lumping tree will differ from v1.0.0 results. Saved Z matrices remain valid for loading, but recomputing from scratch may yield a different tree.	Medium
`source` path resolution — Paths are now resolved relative to the config file, not the working directory. Users relying on CWD-relative `source` paths must update their configs.	Low
`sklearn` now required — `silhouette_score` and `calinski_harabasz_score` require scikit-learn.	Low

Python 3.12/3.13/3.14 CI support; prettypyplot added as dependency.
All tests reclassified as CORE; stochastic tests now use seeded RNG and are fully deterministic.
Tutorial notebook added (docs/tutorial/).
Documentation expanded: quality metrics page, stochastic workflow guide, plot types reference, implied timescales math.