fix(gateway): 避免SSE delta将缓存创建明细重置为0

- 仅在 delta 中 5m/1h 值大于0时覆盖 usage 明细
- 新增回归测试覆盖 delta 默认 0 不应覆盖 message_start 非零值
- 迁移 054 在删除 legacy 字段前追加一次回填，避免升级实例丢失历史写入

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

backend/internal/service/gateway_service.go

@@ -4474,8 +4474,10 @@ func (s *GatewayService) parseSSEUsage(data string, usage *ClaudeUsage) {
 		// 解析嵌套的 cache_creation 对象中的 5m/1h 明细
 		cc5m := gjson.Get(data, "usage.cache_creation.ephemeral_5m_input_tokens")
 		cc1h := gjson.Get(data, "usage.cache_creation.ephemeral_1h_input_tokens")
-		if cc5m.Exists() || cc1h.Exists() {
+		if cc5m.Exists() && cc5m.Int() > 0 {
 			usage.CacheCreation5mTokens = int(cc5m.Int())
+		}
+		if cc1h.Exists() && cc1h.Int() > 0 {
 			usage.CacheCreation1hTokens = int(cc1h.Int())
 		}
 	}

backend/internal/service/gateway_streaming_test.go

@@ -79,6 +79,22 @@ func TestParseSSEUsage_DeltaOverwritesWithNonZero(t *testing.T) {
 	require.Equal(t, 60, usage.CacheReadInputTokens)
 }
 
+func TestParseSSEUsage_DeltaDoesNotResetCacheCreationBreakdown(t *testing.T) {
+	svc := newMinimalGatewayService()
+	usage := &ClaudeUsage{}
+
+	// 先在 message_start 中写入非零 5m/1h 明细
+	svc.parseSSEUsage(`{"type":"message_start","message":{"usage":{"input_tokens":100,"cache_creation":{"ephemeral_5m_input_tokens":30,"ephemeral_1h_input_tokens":70}}}}`, usage)
+	require.Equal(t, 30, usage.CacheCreation5mTokens)
+	require.Equal(t, 70, usage.CacheCreation1hTokens)
+
+	// 后续 delta 带默认 0，不应覆盖已有非零值
+	svc.parseSSEUsage(`{"type":"message_delta","usage":{"output_tokens":12,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0}}}`, usage)
+	require.Equal(t, 30, usage.CacheCreation5mTokens, "delta 的 0 值不应重置 5m 明细")
+	require.Equal(t, 70, usage.CacheCreation1hTokens, "delta 的 0 值不应重置 1h 明细")
+	require.Equal(t, 12, usage.OutputTokens)
+}
+
 func TestParseSSEUsage_InvalidJSON(t *testing.T) {
 	svc := newMinimalGatewayService()
 	usage := &ClaudeUsage{}

backend/migrations/054_drop_legacy_cache_columns.sql

@@ -8,7 +8,37 @@
 --   cache_creation_1h_tokens  (defined in 001_init.sql)
 --
 -- Migration 009 already copied data from legacy → canonical columns.
--- This migration drops the legacy columns to avoid confusion.
+-- But upgraded instances may still have post-009 writes in legacy columns.
+-- Backfill once more before dropping to prevent data loss.
+
+DO $$
+BEGIN
+    IF EXISTS (
+        SELECT 1
+        FROM information_schema.columns
+        WHERE table_schema = 'public'
+          AND table_name = 'usage_logs'
+          AND column_name = 'cache_creation5m_tokens'
+    ) THEN
+        UPDATE usage_logs
+        SET cache_creation_5m_tokens = cache_creation5m_tokens
+        WHERE cache_creation_5m_tokens = 0
+          AND cache_creation5m_tokens <> 0;
+    END IF;
+
+    IF EXISTS (
+        SELECT 1
+        FROM information_schema.columns
+        WHERE table_schema = 'public'
+          AND table_name = 'usage_logs'
+          AND column_name = 'cache_creation1h_tokens'
+    ) THEN
+        UPDATE usage_logs
+        SET cache_creation_1h_tokens = cache_creation1h_tokens
+        WHERE cache_creation_1h_tokens = 0
+          AND cache_creation1h_tokens <> 0;
+    END IF;
+END $$;
 
 ALTER TABLE usage_logs DROP COLUMN IF EXISTS cache_creation5m_tokens;
 ALTER TABLE usage_logs DROP COLUMN IF EXISTS cache_creation1h_tokens;