fix(agent,claw402): harden agent runtime and strip max_tokens for thinking models

- Fix Stop() race condition using sync.Once
- Add ensureHistory() to prevent nil panic in planner/dispatcher
- Add bounds check on trader ID slicing
- Log saveExecutionState and clearSetupState errors instead of discarding
- Remove always-true modelID condition in onboard setup
- Add Chinese setup keywords and expand model name aliases
- Strip max_tokens from claw402 requests to avoid thinking-model budget exhaustion
- Hide Agent nav tab (Beta) pending merge to main
- Sync tests with code changes

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

agent/config_tools_test.go

@@ -6,7 +6,6 @@ import (
 	"strings"
 	"testing"
 
-	"nofx/mcp"
 	"nofx/store"
 )
 
@@ -380,7 +379,9 @@ func TestLoadAIClientFromStoreUserUsesUserSpecificEnabledModel(t *testing.T) {
 		t.Fatalf("unexpected model name: %s", modelName)
 	}
 
-	if _, ok := client.(*mcp.Client); !ok {
-		t.Fatalf("expected *mcp.Client, got %T", client)
+	// After the provider registry refactor, registered providers (like openai)
+	// return their own AIClient implementation, not *mcp.Client.
+	if client == nil {
+		t.Fatal("expected non-nil AI client from provider registry")
 	}
 }