The classifier re-scanned the source on every session; on cloud-backed mounts
(OneDrive/Samba) that's minutes of per-op latency. Workspaces fix it: scan a
folder ONCE, snapshot the completed tree, and resume instantly — all
classification runs on the data model; the filesystem is only touched at copy.
- persist.js v2: multi-workspace IndexedDB (tiny 'index' store for the welcome
list + 'data' store holding the source handle, tree snapshot, and map). DB v2.
- scanner.js: snapshotTree()/loadSnapshot() (compact, handle-less, marked done,
totals recomputed) + lazy resolveFileHandle/resolveDirHandle from the root.
- workspace.js: welcome manager (new/open/rename/delete), debounced autosave of
the active workspace, 'Refresh from disk' (re-scan → re-snapshot, path-keyed
map carries over). New workspace = the one slow full scan; reopen = instant.
- copy.js: resolves snapshot files' handles from the workspace root with a
one-click read permission re-grant; missing-on-disk files surface as errors.
- app.js: enterAppShell() shared by rename/workspace flows; exposes setMode;
classify.js decoupled from persistence.
- template/css: welcome workspace list + header 'Workspaces' button.
- tests: snapshot round-trip, persist CRUD + classify-only-preserves-tree,
copy-from-snapshot via mock root handle (28 classify/classifier tests green).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
