v3.2.0

3.2.0 (2026-09-21)

Bug Fixes

• embeddings: address PR review comments on spec/code inconsistencies (7beddd3)
• embeddings: match provider prefixes case-insensitively (ceb51ef)
• embeddings: recognize namespaced quantized models (fb43b8b), closes #484
• embeddings: resolve provider prefixes against a known provider list (232c81b)
• embeddings: restore the safe credential fallback (a5f0173)
• embeddings: withdraw the unimplemented sagemaker provider (d91d8a2)
• github: key auth error messages off the response status (46625cf)
• scraper: address review findings on outcome and persistence handling (69d8227), closes #490
• scraper: close data-loss paths found auditing merged work (8d38c66)
• scraper: define scrape progress counter semantics (eb7ee20), closes #490 #490
• scraper: detect misfiled script types and cache compiled patterns (42b4c6a), closes #490
• scraper: measure the main-region ratio on normalized text (81acf0a)
• scraper: name the content type when an unprocessable root fails (211934c)
• scraper: record a markdown variant under the page it represents (ce5220b)
• scraper: record the retrieval location of an empty page (b12625e)
• scraper: salvage markdown node by node when conversion fails (e07296c)
• scraper: scope html extraction to the declared main region (2426a04), closes #504
• scraper: stop indexing raw HTML when markdown conversion fails (a338dd0), closes #496
• scraper: stop lowercasing URLs, and specify normalization (98eee82)
• scraper: treat a plain-text markdown alternate as markdown again (0d8b2ba)
• search-eval: fail loudly on a dead Context7 library instead of scoring zero (7afc239)
• search-eval: stop the embedding model spec disabling the regression gate (75c21e1)
• splitter: drop heading anchor annotations from indexed text (4d8bd56)
• splitter: index real markdown frontmatter instead of "undefined" (cc275cb), closes #503
• splitter: only treat a YAML mapping as markdown frontmatter (dde63d1)
• store: address the second review round on page identity (dcb016a)
• store: apply the version label contract on read as well as write (c15279e)
• store: gate version coercion to partial-version shapes (ab939ab)
• store: handle compaction size races (ee0cd03)
• store: keep the location a page's content came from (7ac750f)
• store: keep version labels distinct when normalizing for matching (1b72fb0)
• store: reclaim sqlite space after documentation removal (044d881), closes #476
• store: reduce vacuum memory pressure (6eba99c)
• store: resolve competing page representations without losing content (efd8020)
• store: resolve partial semver versions like "1.20" in findBestVersion (928c424), closes #475 #475 #480
• test: derive the playwright test config from the real defaults (24d0151)
• test: make npm run test:live actually run the live suite (3b592ab)

Features

• scraper: harden fetch fallbacks and llms.txt parsing (77a5648), closes #488
• store: unify version label handling across write and read paths (abaf88e), closes #475 #480 #475 #480

Performance Improvements

• scraper: skip fetching content no pipeline can process (8f42d7a), closes #490