fix: context propagation in LLM calls, SSE error handling, security hardening

- Propagate request context through MCP client (Request.Ctx field)
  — Client disconnects now cancel in-flight LLM API calls (both streaming and non-streaming)
  — Prevents wasted LLM tokens when user navigates away mid-response
- proxyBinance: use http.NewRequestWithContext for context-aware upstream calls
  — Client disconnects cancel Binance proxy requests
  — Distinguishes client cancellation from upstream failures
- Fix SSE parse bug in AgentChatPage: catch block was swallowing error events
  — throw new Error(data) for 'error' events was caught by the same try/catch
  — Now parse JSON separately, then handle events outside try/catch
- Add io.LimitReader to Hyperliquid coins.go JSON decoder (4MB limit)
- Use safe.ReadAllLimited in batch ticker handler for consistency
- Sanitize URL validation error in handler_ai_model.go (was leaking raw error)
- Cache agent tool definitions (built once, reused per message)
- Sort trade history by exit time for consistent ordering

provider/hyperliquid/coins.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"io"
 	"net/http"
 	"nofx/logger"
 	"sort"
@@ -82,8 +83,9 @@ func (p *CoinProvider) fetchCoins(ctx context.Context) error {
 	}
 
 	// Response is an array: [meta, [assetCtxs...]]
+	// Limit read to 4MB to prevent memory exhaustion from oversized responses
 	var rawResp []json.RawMessage
-	if err := json.NewDecoder(resp.Body).Decode(&rawResp); err != nil {
+	if err := json.NewDecoder(io.LimitReader(resp.Body, 4*1024*1024)).Decode(&rawResp); err != nil {
 		return fmt.Errorf("failed to decode response: %w", err)
 	}