perf(dashboard): replace pause/jobs/pool polling with push events

[javi11]

Summary

Three dashboard endpoints were polled by every client on tight intervals:

Endpoint	Poll	Caller
/api/processor/paused (+ auto-paused, reason)	1000 ms	DashboardHeader.svelte
/api/running-job-details	500 ms	ProgressSection.svelte
/api/metrics/nntp-pool	5000 ms	ProviderStatus.svelte, routes/metrics/+page.svelte

That's ~3 req/s per idle client. It also caused the "HTML instead of JSON" symptom in dev: when the Go backend on :8080 was unreachable, the Vite proxy returned the SPA index.html, which crashed JSON.parse in web-client.ts.

This PR replaces the polling with push events over the existing WebSocket hub. No new transport — SSE was considered and rejected in favor of reusing the WS infrastructure already used for upload/config events.

What changed

Backend (internal/backend/)

• app.go — added a ProcessingPauseState payload and pauseState() helper; added App.emit(event, payload) that routes to Wails runtime (desktop) or webEventEmitter (web). PauseProcessing/ResumeProcessing now broadcast the full pause snapshot so the UI doesn't need follow-up RPCs.
• broadcaster.go (new) — one shared ticker per metric:
  • 1 s → running-jobs-updated (JSON dedupe; idle state goes silent)
  • 1 s → processing:auto-paused (state-change detection only)
  • 5 s → nntp-pool-metrics-updated
  • Wired into Startup / Shutdown.

Frontend

• lib/api/events.ts (new) — typed event payloads + event-name constants.
• DashboardHeader.svelte, ProgressSection.svelte, ProviderStatus.svelte, routes/metrics/+page.svelte — drop setInterval loops and visibility-change handlers, fetch once on mount, then subscribe via apiClient.on().

Net effect

Replaces N pollers × per-client with 1 ticker × server-side. A typical idle dashboard goes from ~3 req/s to zero recurring HTTP traffic; updates arrive as WS frames.

Reviewer notes

• The REST endpoints stay — they're still used for the initial fetch on mount. Removing them would force a "wait for first event" race on every page load.
• nntp-pool-metrics-updated is not deduped because Timestamp / Elapsed / AvgSpeed drift every tick; the cost of an unconditional 5 s emit is negligible.
• running-jobs-updated dedupe is most effective when idle (no jobs → identical empty payload → no emit).
• Auto-pause emission piggybacks on the running-jobs tick rather than instrumenting internal/processor/ to avoid an import cycle into internal/backend/.

Test plan

• go build ./internal/... ./cmd/web/...
• go vet ./internal/... ./cmd/web/...
• go test -race ./internal/processor/... ./internal/queue/...
• bun run check — 0 errors, 0 warnings
• Manual: go run ./cmd/web + bun run dev, confirm DevTools shows each of the three endpoints called once on mount and never again; WS frames carry the new event types.
• Manual: toggle pause from the dashboard, confirm header updates with no network round-trip.
• Manual: kill backend then restart, confirm WS reconnect resyncs dashboard state.
• Manual: same flow in Wails desktop (wails dev).