slowave: v0.9.3

0.9.3 (2026-07-04)

Bug Fixes

• both paths now explicitly guard against missing embeddings: (7d9a1e0)
• contradiction judge gates on support count and recency (f20a267)
• contrastive TF-IDF uses global schema corpus as background (87bfdec)
• dashboard caches engine instead of creating per request (0b34f75)
• differentiate CA3/CA1 dual-scale prototype thresholds (c62cf60)
• graph transition/coactivation weights accumulate across replays (abbf206)
• group A non-functional/benchmark-neutral fixes (Homebrew SHA, deps, README benchmarks, license, conda, dead code, docs, gitignore, results) (acbb58e)
• group A non-functional/benchmark-neutral fixes (Homebrew SHA, deps, README benchmarks, license, conda, dead code, docs, gitignore, results) (d525909)
• missing embedding must not trigger supersession (Group B-1) (7d9a1e0)
• persist FAISS indexes to disk via faiss.write_index/read_index (ca2d957)
• profile-layer memories must not be geometry-superseded (Group B-2) (6ad3540)
• remove LLM-era columns from consolidation_debug (e5e94e2)
• scope-filter FTS and prototype candidates at collection time (691cfa1)
• session resolver uses per-thread bindings to prevent collisions (7c0bfe1)
• supersession exception handling — log instead of silently pass (b37b5e0)
• untrack generated wiki benchmark result JSONs (0150f88)

Documentation

• explain transition edges as hippocampal replay co-activation (1823957)
• note that LoCoMo category breakdown is from prior run (3f20fdf)
• update LoCoMo benchmark to 74.6% (latest full suite run) (432b585)