[2.x] fix(realtime): recover desktop tabs after silent socket loss and catch up missed events

[ekumanov]

Fixes #4717.

Full analysis: see the issue. Short version: desktop Safari suspends hidden pages exactly like iOS WebKit, so the WebSocket dies silently while the tab is backgrounded — but #4662 gated the #4590/#4654 recovery to isIOS(), pusher-js's own reconnects perform no catch-up on any platform, and a ≥2-post gap permanently disables live-append in an open discussion because PostStreamState.update() is guarded by viewingEnd()'s 1-post drift tolerance. Net effect observed in production: a missed reply never appears (quiet discussion), or appears ~50 minutes late only because newer replies happened to force a range load.

Changes

extensions/realtime — extend/Application.ts

• The visibilitychange recovery now fires for isIOS() (unchanged) or whenever the connection is demonstrably unhealthy: connection.state !== 'connected', or no protocol frame received within the activity window (65 s — server-advertised activity_timeout 30 s + pong wait), tracked via a bind_global last-activity timestamp. Healthy desktop tabs keep receiving pusher:pong every ≤30 s even while hidden (verified in Chrome with a 15-minute hidden-tab soak), so plain tab switches still cause no spurious refetch — the [2.x] fix(realtime): only force-reconnect on iOS visibilitychange #4662 complaint stays fixed; the zombie case (state says connected, socket long dead) now reconnects.
• New watchConnection() applied to the initial Pusher instance and to every fresh instance forceReconnect() builds: on any transition into connected after a previously established connection — pusher-js-internal recovery as well as forced rebuilds — it runs catchUp(). This replaces the one-shot onReconnected refresh that only the iOS path used to get.
• catchUp() refreshes the discussion list (as before) and re-syncs an open DiscussionPage: it captures stream.viewingEnd() before the refetch (reliable at that point — the missed posts haven't reached the store yet), then store.find('discussions', id) and, if the user was at the end, loads through to the new count.

extensions/realtime — extend/Discussion/NewActivity.ts

• Both websocket handlers now capture viewingEnd() before app.store.pushPayload(...) updates postIds, and call stream.syncEnd() when the user was at the end. Previously, one missed event put the drift at ≥2 and every subsequent update() returned early — live-append was permanently dead for that page view even though events were arriving fine.

framework/core — states/PostStreamState.ts

• New syncEnd(): load through to the latest post, expanding the visible window — i.e. update() without the drift guard, for callers that have already established end-ness. update() is refactored to delegate to it and documents the caveat. No behavior change for existing callers.

Testing

• framework/core/js/tests/unit/forum/states/PostStreamState.test.ts: three new cases — update() appends a single new post when viewing the end; update() refuses at ≥2-post drift (locks the guard semantics the fix relies on); syncEnd() loads through regardless of drift. All 10 cases in the file pass.
• The realtime extension has no JS test harness, so the reconnect logic is covered by manual verification: foreground delivery, hidden-tab delivery, and reconnect catch-up exercised against a production rc.3 forum (self-hosted websocket server behind Cloudflare) with two accounts in a private discussion.
• yarn build and yarn check-typings pass for the extension; yarn check-typings passes for core.

Per convention, no js/dist/dist-typings are included.