What's new
Features
- FoundationEncoder — new latent algorithm for foundation model embeddings
- MismatchRatio metric + relative PersistentHomology threshold
- RobustPCA — global (ADMM/IALM) + local (MCD/trimmed/Huber/Gaussian) + LTSA variants
- BalancedLabelSampling and DiffusionCondensationSampling samplers
- Algorithm registry —
get_algorithm()andlist_algorithms()for programmatic discovery - Hydra-free API —
api.pynow uses Python registries instead of Hydra compose
Fixes
- Correct t-SNE perplexity formula (was
neighborhood_size * 3, nowneighborhood_size) - Pre-fit sampling label alignment and label-aware sampler context
- Seed/neighborhood_size forwarding in registry path
- experiment.py wandb, scores overwrite, imports, and types
Refactors
- Split
experiment.pyinto Hydra-free engine (run_experiment) - Remove Hydra from
api.py— pure importlib instantiation - Unified metric routing and sampling with
at:field
Docs
- Added
ARCHITECTURE.mdwith codebase map, data flow, and invariants - Trimmed
CLAUDE.mdto behavioral rules only
Full changelog: v0.1.3...v0.1.4