feat(wave3): missing actions + error visibility

[vikranthreddimasu]

Context

Wave 3 of 5 from PLAN.md. Wave 1 (#12) stopped the bleeding; Wave 2 (#13) resurrected the citation layer. This one closes table-stakes gaps: rename, delete-document, empty-notebook creation, error persistence, and Ollama health monitoring.

Actions that didn't exist

• Rename notebook. Backend `PATCH /notebooks/{id}` + `notebook_store.rename_notebook`. Frontend: context-menu "Rename" → inline edit, same pattern as conversations. New notebook creation now drops you directly into rename mode so "Untitled notebook" doesn't stick around.
• Delete document. Backend `DELETE /documents?notebook_id=&source_path=` removes chunks from the chroma collection, the summary row, the file on disk (scoped to the notebook's `uploads_dir` with the same `Path.relative_to` guard as preview), and decrements the notebook's counts via a new clamp-at-zero `adjust_counts` helper. Frontend: overflow menu on each document card with confirmation.
• Create empty notebook. `+ New` now calls `POST /notebooks/` directly instead of forcing an immediate file upload.
• Zotero is discoverable. Button lives next to `+ Add` in the Documents section — previously only reachable via the command palette.

Error visibility

• Notification center. Bell icon in the sidebar footer with an unread badge; opens a slide-over panel listing the last 50 errors, successes, and warnings with relative + absolute timestamps. `Toast.tsx` now mirrors non-`info` toasts into this log so a failure at second 3 is still recoverable at second 30.
• Human error messages. New `utils/errorMessages.ts` translates common failures:
  • `ECONNREFUSED` / `Failed to fetch` → "The backend isn't reachable. Check that the server is running."
  • 413 → "That file is too large. Split it into smaller documents."
  • 415 → "That file type isn't supported. Try a PDF, DOCX, PPTX, TXT, or Markdown file."
  • 422 → "We couldn't process that. The file may be empty, scanned without OCR, or corrupt."
  • 403 → "That file is outside this notebook."
  • Ollama / model-not-found → "The AI model isn't available. Make sure Ollama is running."
  • Unknown errors fall through to the raw message (capped).
• `role="alert"` on error toasts so screen readers announce immediately.
• AppShell drop handler no longer aborts the whole batch. Each file has its own try/catch — file 2 failing doesn't swallow the result of file 1, and the error message names which file failed.

Reliability

• `ConnectionBanner` monitors Ollama too. Three-state: `healthy` / `backend-down` / `ollama-down`. Ollama crash mid-session now shows an amber warn-banner: "AI model (Ollama) isn't reachable. Start Ollama, then try your message again." Previously a dead model surfaced only as a raw error in the chat bubble.
• Banner runs its check on mount instead of only after the first 30s interval tick. Interval dropped from 30s to 15s.

Verified

• `cd apps/desktop && npm run build` — clean (272 modules, 847ms)
• `cd backend && uv run ruff check .` — clean
• `cd backend && uv run pytest -q` — 33 passed (3 new: rename happy path, rename-missing returns None, `adjust_counts` clamps to zero)

Intentionally not in this PR (punted to Wave 3b or Wave 4)

• 3.3 duplicate upload detection — needs ingestion-layer change; belongs with Wave 4's other vector_store work
• 3.4 real upload progress bar — requires an XHR refactor of the upload path; kept separate for review surface
• 3.6 wizard model selection wiring — needs a backend model-override API; scoped out here

🤖 Generated with Claude Code