Notes from the repositories.

The collector builds an authorized source plan from configured caches, repository inventory, reader selection, and per-repository modes. A separate selection policy ranks qualifying findings, fits them to a reading budget, and publishes included, deferred, excluded, quiet, and inaccessible decisions with the edition.

Select one cached repository and one uncached repository, export inventory, collect, then confirm the cached repository is inspected, the uncached repository is listed as inaccessible, and every published story is accounted for in the selection ledger.

The exporter currently fetches feedback and repository inventory in one Promise.all; if the inventory function is not deployed, the export aborts before the collector can use its intentional legacy-cache fallback.

Topic threads and follow-up questions now retain repository, base and head commits, path, and bounded lines. Active watches and queued questions feed the compiler, archived editions form a filterable history, and the Continue planner ranks unread stories, re-reviews, map areas, and open questions into the reader's chosen attention window.

Watch one code excerpt, queue a question on its exact lines, mark another map area for revisit, then reload and confirm Continue contains all three items within the selected time window and History can find the original topic.

Continue priorities and minute estimates are hard-coded heuristics, so the queue is deterministic but not yet calibrated from observed reading behavior; low-ranked items may remain unseen in short windows.

The extension gained tab, URL-list, bookmark, and multi-image X intake with persisted checkpoints and a shared captureSessionId. Owner routes now require the deployment key, while a ten-minute one-use grant gives the Clipper a hashed, revocable device credential limited to capture.

Pair one Clipper, start a three-item batch with one bad URL, reload and resume, confirm two references share the session, then revoke the device and verify its next capture is rejected.

The vault remains a single-owner bearer-secret model, and session summaries are inferred from recent successful references, so failed or skipped items and original expected counts are not reconstructed.

Ourchival added boards, tags, projectReferences with per-use notes, bounded batch actions, and saved searches with field-level match excerpts. A separate enrichment ledger produces pending tag suggestions and reason-bearing related-reference scores rather than mutating taxonomy silently.

Attach two references to a project with a reason, save a search for it, accept one suggested tag, then confirm search and Related references show labeled match or ranking reasons.

Retrieval remains bounded application-layer scanning and enrichment uses hand-tuned English rules; older rare matches can require paging, and the scores should not be read as learned semantic relevance.

A scheduled Node action now claims media_derivatives jobs, loads a private Drive or Convex original, enforces byte and pixel limits, generates two oriented WebP sizes, computes content and perceptual hashes plus dominant colors, stores the outputs, and atomically marks the asset and job ready.

Enqueue one stored image without derivatives, wait for the worker, then confirm the asset has previewStorageId, thumbStorageId, dimensions, SHA-256 contentHash, 64-bit perceptualHash, colors, and derivativeStatus ready while the original provider fields remain unchanged.

Only stored originals are eligible, failed attempts are not automatically retried without a forced enqueue, and a failure after writing one or both blobs but before the completion mutation can leave unreferenced storage objects for later cleanup.