New saves can join an accepted topic and area through the existing connection request. Placement needs a known catalog ID, confidence of at least 0.90, a same-subject decision, an explanation and literal source evidence. Accepted groups and concurrent manual choices stay protected, including exact inheritance of a manually chosen topic without an area. Uncertain matches retain the extracted subject. An accepted placement can be inspected and its original topic restored.
Text-only captures now use numbered source passages for concept evidence in both Claude and Gemini. Photo/OCR extraction retains its quotation path. Extraction remains source-only; library placement and user-reviewed concept aliases happen afterward.
Manage concepts lets you explicitly merge equivalent names, reuse aliases on future captures, and Undo after reload or JSON backup. Undo protects later concept/source edits. Connection retrieval considers captured source text and reviewed aliases; passage selection covers long unpunctuated text and Unicode matches. Existing direct-evidence and generic-link guards remain in force.
Topic grouping works in batches of up to 150 pieces, carries topic/area labels between batches, and saves completed work in IndexedDB. Resume validates the snapshot and does not replay completed batches. Apply still requires an explicit preview choice, with existing whole-group stale checks and Undo. Changed snapshots require discarding saved batches.
Library health shows processing exceptions and coverage counts, with links to affected pieces and existing review tools. Record health check saves up to 12 local aggregate snapshots in JSON backups. A singleton topic is not an error, and these counts are not a semantic quality score.
Validation includes code tests, a synthetic guardrail benchmark, interrupted grouping of 321 pieces, phone/desktop growth flows, a complete 320px browser regression pass, offline assets and a private backup invariance check. The full Chromium/WebKit matrix must pass before merge. No real provider calls or private-content commits were used for validation.
New-save placement shares a request but adds some input/output tokens. Bulk grouping and reviews remain explicit paid actions. Candidate retrieval is lexical, with bounded context; placement considers up to 60 accepted topic candidates. Semantic correctness still depends on model judgment and user review. Existing pieces are not automatically reclassified or reprocessed by this update.