[codelion]

Summary

Version 0.3.1. Follow-up to #68, from a systematic hunt for bugs in features that had little or no test coverage: lifecycle operations, multi-label, strategic mode, the HTTP server, the scikit-learn wrapper, merge_classifiers, and the CI/release plumbing. Three independent reviewers (core library, feature modules, CI/packaging/docs) read the code cold and tried to break it; every finding below was reproduced with a script before being fixed. CHANGELOG.md has the full list.

Core library

• add_examples is now all-or-nothing. An error or Ctrl-C while training used to leave the label map ahead of the neural head, so every later predict raised "selected index k out of range" and every retry failed; or left the search index stale, so predictions silently returned the wrong class. Texts are encoded before any state changes, and on failure everything is restored.
• EWC did nothing. The penalty compared the old head with its own stored copy, so it was always exactly 0 and never reached the head being trained. It now penalises how far the live head's existing rows move. With the default strength (5.0) the effect is small: re-running the README's class-incremental experiment (MiniLM, ag_news and emotion, 3 seeds) moved final accuracy by -0.004 / +0.016, old-class loss by -0.002 / +0.002 and new-class recall by -0.012 / 0.000, all within noise. The README table therefore stands.
• merge_classifiers was badly broken: the other classifier's classes were never searchable afterwards (the prototype index was not rebuilt; 0% accuracy on them), merging into an empty classifier raised, training counts were not combined, stale calibration was kept. Also fixed a deadlock when two threads merge each other (it never holds two classifiers' locks at once) and the other classifier's OOD radii are kept.
• clear_memory() with no labels cleared only the prototypes; the label map, counts, calibration and neural head stayed, so predict kept answering from the head. It now empties the classifier.
• A reloaded classifier re-averaged its prototype from the few kept examples on the first add_examples (moved by 0.32 versus 0.03 live for a 30-example class). Prototypes now fold in new examples with the weight of the examples they were saved from.
• Re-saving an ONNX-loaded classifier wrote the ONNX folder as model_name; add_examples ignored batch_size; copy.deepcopy/pickle failed because of the instance lock; predict_batch disagreed with predict in strategic and multi-label modes; blank training texts/labels were silently learned (behaviour change: they now raise ValueError); remove_examples("str") removed nothing; PrototypeMemory.get_strategic_prototypes raised NameError.

Multi-label, strategic mode, server, scikit-learn

• Multi-label: forget/remove_examples/load(use_onnx=...) crashed; an explicit threshold was ignored; a saved multi-label model loaded through AdaptiveClassifier.load (or the server) came back as single-label; the head was barely trained; calibrate/predict_set returned meaningless numbers (now NotImplementedError); bare-string labels were split into one class per character.
• Strategic mode: the README's example config silently disabled it (costs must be a list with one entry per embedding dimension; the reason is now logged); the best response only explored the first few dimensions and was unseeded; scores depended on k; the robustness evaluation divided by zero and crashed unless the levels included 0.0 and 1.0.
• Server: saves, updates and /info no longer block the event loop (a save with ONNX export froze /health for seconds); multi-label models are refused with a clear message instead of failing every request with 500.
• scikit-learn: partial_fit(classes=...) now lists declared classes; labels that collapse to the same class name (1 and "1"), None and NaN are rejected.

New

• examples/javascript/portable_inference.mjs: a Node.js port of the portable inference, tested against predict() for both pooling modes, every scoring branch and over-length inputs (tests/test_js_example.py, run in CI).
• scripts/docker_smoke.py and .github/workflows/docker.yml: build the serving image and query it (path-filtered plus weekly, so ordinary PRs do not build it).
• Release workflow now calls the full test workflow, only the publish job holds PyPI credentials, and the release tag must match the package version.
• Docs: docs/API.md now covers the real signatures, MultiLabelAdaptiveClassifier, strategic mode and what a saved classifier keeps; README documents the new-class recall trade-off (new_class_example_threshold) with measurements.

Testing

• 95 new test functions (more cases once parametrised): state safety, merge, multi-label, strategic mode, server, scikit-learn, Node port, memory. Most were checked to fail against the previous code.
• The full offline suite passes locally (0 failures; two SetFit skips). Tests that download real models could not run in the authoring sandbox (Hugging Face rate-limits its IP), so CI's prefetched run is their first check.

Known gaps

• Unverified until CI runs: the Docker build and smoke test (the sandbox's TLS proxy blocks the PyTorch download inside docker build), the onnxruntime-node path of the JavaScript example (only onnxruntime-web could be installed locally), and the reworked release workflow.
• A reloaded classifier's neural head is still retrained from the few examples kept per class (documented in docs/API.md); only the prototypes were fixed.
• Strategic mode is still not covered by the portable reference or the accuracy/calibration tests.

🤖 Generated with Claude Code