robustness: shared HTTP client, HTTP status checks, abort controller, streaming perf

- web.go: reuse shared binanceClient with connection pooling instead of per-request client
- sentinel.go: check HTTP status code before parsing Binance ticker response
- brain.go: check HTTP status codes in news scan and market brief APIs
- brain.go: add cleanStaleSignals() to prevent unbounded sync.Map growth
- scheduler.go: periodically clean expired pending trades
- AgentChatPage.tsx: add AbortController to cancel in-flight requests
- AgentChatPage.tsx: check response.ok before parsing JSON
- AgentChatPage.tsx: batch word streaming (~40 frames max) to reduce re-renders
- AgentChatPage.tsx: handle AbortError gracefully (remove orphan bot message)

agent/scheduler.go

@@ -43,10 +43,14 @@ func (s *Scheduler) Start(ctx context.Context) {
 					lastCheck = now
 				}
 				// Clean stale chat history every hour (sessions idle > 24h)
+				// and expired pending trades
 				if now.Sub(lastCleanup) > 1*time.Hour {
 					if s.agent.history != nil {
 						s.agent.history.CleanOld(24 * time.Hour)
 					}
+					if s.agent.pending != nil {
+						s.agent.pending.CleanExpired()
+					}
 					lastCleanup = now
 				}
 			}