fix(client): support vLLM "reasoning" field for thinking blocks

vLLM sends thinking content in a "reasoning" delta field, unlike
DeepSeek which uses "reasoning_content". Check both field names so
thinking blocks render for vLLM-hosted models like qwen3.6-27b-thinking.

Also update client tests to exercise thinking output and skip by default
so they don't run in Drone CI (require live LLM API).

backend/internal/client/client.go

@@ -115,15 +115,19 @@ func (c *Client) SendMessage(ctx context.Context, chatMessages []*store.Message,
 		if len(chunk.Choices) > 0 {
 			delta := chunk.Choices[0].Delta
 
-			// Check Thinking
-			if thinkingField, found := delta.JSON.ExtraFields["reasoning_content"]; found {
-				var thinkingContent string
-				if err := json.Unmarshal([]byte(thinkingField.Raw()), &thinkingContent); err != nil {
-					return respContent, fmt.Errorf("thinking unmarshal error: %w", err)
-				} else if thinkingContent != "" {
-					msgStats.RecordFirstToken()
-					sendUpdate = true
-					msgChunk.Thinking = ptr.Of(thinkingContent)
+			// Check Thinking - Support both "reasoning_content" (DeepSeek)
+			// and "reasoning" (vLLM) field names.
+			for _, thinkingKey := range []string{"reasoning_content", "reasoning"} {
+				if thinkingField, found := delta.JSON.ExtraFields[thinkingKey]; found {
+					var thinkingContent string
+					if err := json.Unmarshal([]byte(thinkingField.Raw()), &thinkingContent); err != nil {
+						return respContent, fmt.Errorf("thinking unmarshal error: %w", err)
+					} else if thinkingContent != "" {
+						msgStats.RecordFirstToken()
+						sendUpdate = true
+						msgChunk.Thinking = ptr.Of(thinkingContent)
+					}
+					break
 				}
 			}