perf(relay): full-chain optimization — key pool, model sync, SSE stream

Phase 1 (Key Pool correctness):
- RPM: fixed-minute window → sliding 60s aggregation (prevents 2x burst)
- Remove fallback-to-provider-key bypass when all keys rate-limited
- SSE semaphore: 16→64 permits, cleanup delay 60s→5s
- Default 429 cooldown: 5min→60s (better for Coding Plan quotas)
- Expire old key_usage_window rows on record

Phase 2 (Frontend model sync):
- currentModel empty-string fallback to glm-4-flash-250414 in relay client
- Merge duplicate listModels() calls in connectionStore SaaS path
- Show ModelSelector in Tauri mode when models available
- Clear currentModel on SaaS logout

Phase 3 (Relay performance):
- Key Pool: DashMap in-memory cache (TTL 5s) for select_best_key
- Cache invalidation on 429 marking

Phase 4 (SSE stream):
- AbortController integration for user-cancelled streams
- SSE parsing: split by event boundaries (\n\n) instead of per-line
- streamStore cancelStream adapts to 0-arg and 1-arg cancel fns

desktop/src/store/saasStore.ts

@@ -425,6 +425,12 @@ export const useSaaSStore = create<SaaSStore>((set, get) => {
       stopTelemetryCollector();
       stopPromptOTASync();
 
+      // Clear currentModel so next connection uses fresh model resolution
+      try {
+        const { useConversationStore } = require('./chat/conversationStore');
+        useConversationStore.getState().setCurrentModel('');
+      } catch { /* non-critical */ }
+
       set({
         isLoggedIn: false,
         account: null,