Publish the existing H3 merge/export and autotuner correctness fixes, offline character and AV calibration tools, provenance audits, and regression coverage. Add a dictionary-preserving storage hook for offline mapped TIES exports.
Keep scoring weights and defaults unchanged, retain NP/CT as opt-in, and exclude the rejected F08 search-admission prototype. Validate the exact staged tree: 1052 Python tests, 69 subtests, 3 JS tests, and real ComfyUI CPU export round trips pass; 3 CUDA-only tests skipped.
Correct conflict-sparsification skips and scoring, preserve experimental precision, add native-factor NP/CT exports, and harden mixed H3/CLIP detection and external evaluation.
Include regression coverage and reproducible local H3 benchmark, review, and waveform diagnostics with evidence records. Keep experimental preference ranking and package version unchanged.
- ensure_index_fresh now rebuilds on ESTIMATOR_INDEX_VERSION mismatch,
so a schema bump actually invalidates the cached pickle.
- Corrupt meta.json rebuilds instead of crashing.
- Index builder no longer falls back to the whole cache dict when
per_prefix is missing — that previously treated top-level keys
(algo_version, created_at) as prefix stats.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Offline pipeline: snapshot_download the HF community dataset, parse
config/pair/lora JSONs into a phase1 dict per combo, featurize via
EstimatorFeatureExtractor, z-score-normalize, fit NearestNeighbors
(cosine), pickle to disk with meta.json sidecar recording the
hf_commit_sha, index version, feature dim, and build timestamp.
Compound arch filenames like `sd_unet` are parsed as a two-token
suffix so the leading token isn't mistakenly captured as a hash.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ethanfel