[ThomsenDrake]

Summary

Cestus can now import a selected source folder, read and search its documents, follow stable passage citations, and run explicitly approved external document processing through the ordinary browser UI. This delivers the implemented Phase 2 evidence foundation on top of the authenticated Phase 1 workspace.

Equal source bytes share an immutable original while preserving occurrences. Changed sources and upgraded extraction preserve older citations. UTF-8 text, CSV, and PDF text extraction are supported; mixed PDFs retain readable pages and explicitly identify unextracted pages. Textless pages are not assumed blank, and OCR is not implemented.

External processing binds approval to exact content, selected passages and coverage, destination/model, operation, and limits. Credentials remain server-side. Definitive rejection or fully received invalid output fails; ambiguous submission remains uncertain. Neither outcome implies unbilled, and potentially billable retries require a fresh approval. Current governance and blob integrity govern reads, snippets, and transfers.

Verification

• Final-code npm run verify: 4,047 tests passed, five skipped; typecheck, both assurance tests, and production build passed. No implementation changed after this run. The branch has no divergence from the fetched neo base.
• Built UI with the production server and disposable synthetic workspace: registration through extraction, text/CSV/PDF reading, content-phrase search, page-3 navigation with page-2 coverage gaps, duplicate occurrences, changed sources, saved citations, and restart recovery demonstrated.
• Real HTTP transport against a separate local synthetic provider: zero requests at preview/approval; one invalid-response failure; a separately approved retry completed once with a cited result. Restart caused no extra submission. Production regressions cover rejected transfers, ambiguous outcomes, cancellation, and stale authorization.
• On 100 synthetic documents (1,240,176 original bytes, 6,000 passages), request-scoped replay reduced measured queries from 14–15 seconds to 235–326 ms. This is modest-corpus evidence, not a large-capacity guarantee.
• CI installs the required Poppler tools before running the full suite on a clean Ubuntu runner.
• Native review found no actionable findings. Existing source-map and build-size/browser-externalization warnings remain.

Acceptance record and screenshots · Hands-on runbook

Acceptance still pending

Real-record format/scan acceptance requires an authorized source-folder path and required formats. A live provider trial requires the intended endpoint/model, server-side configuration, exact authorized selection, and evaluation budget. Synthetic protocol acceptance does not establish live-provider compatibility or billing. Merging this foundation does not mark all of Phase 2 accepted or introduce Phase 3 workflows.