[iLevyTate]

A review sweep over the analysis path after #107 and #108: the Discover run in the renderer, the batch handler, the model memory manager, the vision runtime, Whisper, the smart-folder watcher, and dates. Each fix has a regression test that was run against the code before the fix and failed there for the reason the fix addresses. CHANGELOG.md ([Unreleased] → Fixed) has the user-facing wording for all of it.

What changes for the user

Discover runs

• A run with a long recording is no longer cut off at 30 minutes or taken for stuck after five minutes without a finished file. Audio and video always go through the batch, the batch keeps reporting during one long file, and the run is bounded by its own checks (no progress; six hours at most).
• Stop leaves nothing behind: no progress after the cancel, no "failed" document, no filename-only "Limited analysis" image.
• "Analysis paused" no longer appears after Stop. A progress update held back by the 80 ms throttle was written after the analysis had ended.
• Leaving Discover no longer says "Analysis cancelled by user", and a run left behind no longer writes over the progress of the run that replaced it.
• The Dashboard waits for a running analysis instead of ending it silently (the way back to Guided stays open).
• Progress from Organize batches, model downloads and Reanalyze all no longer shows as analysis progress, and another sender can no longer take over the running batch.
• A file that finishes right after another appears at once instead of with the next file.
• A run that waited for an earlier batch counts the whole run, in the progress ("31 of 40", not "1 of 10") and in the completion notice.
• A second Analyze right after the first is queued behind it. Until a run wrote its first progress it read as a stale lock, and the second request reset it.
• A file's status is no longer put back by a later change: setFileStates(prev => …) got the list the hook last rendered with, not the one in the store.
• Development builds (StrictMode): the run aborted while starting stops there, and no longer switches off the heartbeat of the run that replaced it.

Models and memory

• Memory counts from the moment a load begins, a second caller for the same model shares the load, and background work does not start on memory that is spoken for. On a 6 GB card the text model and the vision runtime could both start "on paper".
• The vision runtime is no longer shut down under an image (idle timer during another image, a text model taking its memory between start and request, a start that failed without reporting it), a timed-out request no longer pins it for the session, and a crash part way through an answer no longer leaves the image waiting.
• Cancelling one analysis no longer cancels another that shares its transcription.
• A recording longer than the transcription limit is recorded once as "not transcribed" instead of being decoded and failed on every attempt.

Smart-folder watcher

• It asks the history again when a file's turn comes, records the file as it was when the analysis started, pauses an image batch when a Discover run starts, and keeps a reanalysis request when the same file changes.

Dates

• Suggested names, the file date given to the model, dates read from the content and the settings export name use the local calendar day, not the UTC day. A date-range search of the history no longer skips the first or last month of the range away from Greenwich.

Logs

• A model load refused because another model is in use is not logged as an unclassified failure; waiting on a model in use warns once.

Tests

New: ModelMemoryManager.pendingLoads, calendarDayLocalTime (each zone in a child process, since CI runs in UTC and Jest copies process.env), fileStatesUpdaterLiveStore, imageAnalysis.cancellation, ipcMiddleware.progressAfterStop, ipcMiddleware.runBase; additions to the existing suites of the files touched.

The child process of calendarDayLocalTime is plain Node, so it does not have Jest's module mapping. One of the modules it loads requires electron, whose entry point downloads the Electron binary when it is not on disk and prints "Downloading Electron binary..." to stdout. That failed the first zone on a fresh Ubuntu runner (first push of this PR). The probe now puts an empty module in place of electron, as the mapping does (second commit).

test/useAnalysis.test.js now clears the module-level "resume after unmount" flag after each test. A test that unmounted in the middle of a run started a resume in the next test; the stale-lock takeover fixed here used to hide that.

Local: typecheck, typecheck:js, format:check, verify:checkjs-ratchet (493/493), eslint on the changed files, and the full Jest suite.

Looked at and left

• A run that was aborted but not replaced still continues past its start-up pause to the next abort check; its finally handles it and it sends nothing (fix(discover): an aborted run sends no batch; the watcher yields to Discover runs #108).
• The watcher retries an embedding failure without backoff, bounded by MAX_ANALYSIS_RETRIES.
• sendOperationProgress throttles through one pending slot shared by all senders, so one sender's held update can replace another's. Two senders at once would take an Organize batch during a Discover batch, which navigation blocks.
• ModelMemoryManager.pendingLoads.test.js uses real timers with 20–70 ms waits inside a 500 ms window.
• The Dashboard gate is a UX choice and easy to reverse: NavigationBar only.

Not verified here: behaviour on real hardware beyond the dev build on one 6 GB laptop GPU, and the open CodeQL alerts.