[Bug] Hosted relay intermittently stalls desktop message loading and disconnects

**Describe the bug**
Buzz Desktop 0.5.4 intermittently takes several seconds to load messages and surfaced a relay-disconnected state on a hosted `*.communities.buzz.xyz` workspace. The Mac and app are otherwise idle. Direct probes isolate the delay to the hosted relay/WebSocket path rather than the local network or renderer CPU.

**Steps to reproduce**
1. Open Buzz Desktop against a hosted `*.communities.buzz.xyz` workspace.
2. Switch to a DM/channel and wait for messages to load.
3. Observe multi-second loading and occasional relay disconnect.
4. Probe the same relay from the same Mac:
- Ten HTTPS relay-information requests: TTFB ranged **0.720–4.705 s**.
- Ten WebSocket REQ probes: connection/query completion ranged **0.778–4.307 s**; two connection errors and one 10 s timeout.
- `/health` remained fast at **0.109 s**.

**Expected behavior**
Existing messages load promptly, WebSocket subscriptions stay connected, and relay responses do not intermittently take multiple seconds or time out.

**Version and platform**
- Buzz Desktop: 0.5.4 (0.5.4), current latest release as of 2026-08-04
- OS: macOS 27.0 (26A5388g), Apple silicon
- Relay: hosted `*.communities.buzz.xyz` (specific workspace omitted because the report is public)

**Logs / additional context**
Evidence captured 2026-08-04 around 10:00 America/Los_Angeles:

- Local network was healthy: base RTT ~40 ms, download throughput ~73.8 Mbps.
- Relay TCP/TLS setup was fast (~20–38 ms TCP, ~36–203 ms TLS); the delay accrued after TLS while waiting for the relay/application response.
- Buzz desktop process was idle (~0% CPU); WebContent was ~0.4–0.8% CPU during sampling.
- Relay `/health` returned 200 in 0.109 s while the NIP-11/root route took 3.066 s in the same probe set.
- App restart did not prevent recurrence; the reported slowdown occurred roughly seven minutes after a fresh 0.5.4 launch.

This points to an intermittent hosted relay/application-path latency or subscription-recovery issue, not a proven desktop rendering bug. Please inspect hosted relay request/subscription latency and desktop reconnect behavior together.