[shipitdev]

The local assistant could drop applicability prose while keeping download tables, and browser hostname normalization rejected its own UI-validation requests. Preserve retrieved opening prose, explicitly reject unsupported claims, and allow one repair followed by verification. Share the browser origin check across asking, UI validation and session clearing.

Add versioned document updates with changed-only embeddings, isolated staging, activation, rollback, matched SQLite/Qdrant backups and recovery. The API reloads activated versions and invalidates old exact answers. Bound retrieval caches, operational logs and opt-in private traces. Publish operating instructions and redacted V1 acceptance measurements.

Validation

• 141 local tests pass with one environment-specific skip and three subtests. A clean public copy passes 126 tests with the same skip. The separate RAGAS environment passes nine focused tests, including that integration. Suites overlap.
• Eight frontend contract tests, typecheck, lint and production build pass.
• Saved 97-question training replay improves context reference-word coverage from 80 to 83, with three gains and zero losses.
• Real-corpus add/update/delete drill embeds two chunks and reuses 99,695 vectors. Repeating the update embeds zero. Activation and rollback pass.
• Native backup restores 28,482 documents and 99,697 vectors. Three exact-search top-ten comparisons match IDs and scores.
• Four verifier canaries pass. Final targeted generation gives one answer, one clarification and two withheld responses. Live work uses 15 HTTP 200 responses and 62,248 reported tokens.
• Browser checks cover GenUI, excerpts, structured metrics, regeneration and session clearing with the existing controlled response fixture. Final local health checks pass.

Scope is a local single-worker V1. Q046/Q037 still withhold. Training/context proxy results and targeted checks do not establish full answer accuracy. Broader traffic and frozen held-out evaluation remain later milestones. Private sources, traces, backups, learning notes and keys stay local.

A final recovery regression check verifies that a missing activation pointer stops serving and that explicit activation restores a validated version.