v0.9.0

What's Changed

• Fix partial cache dedup and consolidate duplicated helpers by @AlliedToasters in #195
• Add tqdm progress bars to load_activations and pull_dataset by @AlliedToasters in #196
• Remove Probe(dataset=), promote load_activations() workflow by @AlliedToasters in #197
• Consolidate shard tensor slicing into shared helpers by @AlliedToasters in #198
• v2 dataset format: embed metadata in Parquet schema by @AlliedToasters in #200
• Consolidation mega-PR: reduce ~794 lines of duplication by @AlliedToasters in #212
• Update docs and tests for Probe rename, remove outdated content by @AlliedToasters in #213
• Consolidate duplicated code across codebase by @AlliedToasters in #214
• Add comprehensive type hints and mypy CI check by @AlliedToasters in #215
• Consolidate test duplication: -564 lines by @AlliedToasters in #217
• Strip plotting module and CLI tool by @AlliedToasters in #219
• Increase test coverage toward 80% by @AlliedToasters in #220
• Deduplicate model caches, last-token index, and SHA-256 hashing by @AlliedToasters in #224
• Add ActivationStore for in-memory cross-fold workflows by @AlliedToasters in #226
• Add SGDGPUClassifier for GPU-accelerated probe training by @AlliedToasters in #225
• Add LR scheduler and early stopping to SGDGPUClassifier by @AlliedToasters in #230
• Support logits-only extraction without requiring layer activations by @AlliedToasters in #232
• Next release: breaking changes + v2 dataset format by @AlliedToasters in #199

Full Changelog: v0.8.9...v0.9.0