fix: 三端联调 P1 修复 — API密钥页崩溃 + 桌面端401恢复 + 用量统计全零

P1-03: vite.config.ts proxy '/api' → '/api/' 加尾部斜杠，
  防止前缀匹配 /api-keys 导致 SPA 路由崩溃

P1-01: kernel_init 增加 api_key 变更检测（token 刷新后自动重连），
  streamStore 增加 401 自动恢复（refresh token → kernel reconnect），
  KernelClient 新增 getConfig() 方法

P1-02: /api/v1/usage 总计改从 billing_usage_quotas 读取
  （authoritative source，SSE 和 JSON 均写入），
  by_model/by_day 仍从 usage_records 读取

crates/zclaw-saas/src/model_config/service.rs

@@ -419,21 +419,33 @@ pub async fn revoke_account_api_key(
 pub async fn get_usage_stats(
     db: &PgPool, account_id: &str, query: &UsageQuery,
 ) -> SaasResult<UsageStats> {
-    // Optional date filters: pass as TEXT with explicit $N::timestamptz SQL cast.
-    // This avoids the sqlx NULL-without-type-OID problem — PG's ::timestamptz
-    // gives a typed NULL even when sqlx sends an untyped NULL.
+    // === Totals: from billing_usage_quotas (authoritative source) ===
+    // billing_usage_quotas is written to on every relay request (both JSON and SSE),
+    // whereas usage_records has 0 tokens for SSE requests. Use billing as the primary source.
+    let billing_row = sqlx::query(
+        "SELECT COALESCE(SUM(input_tokens), 0)::bigint,
+                COALESCE(SUM(output_tokens), 0)::bigint,
+                COALESCE(SUM(relay_requests), 0)::bigint
+         FROM billing_usage_quotas WHERE account_id = $1"
+    )
+    .bind(account_id)
+    .fetch_one(db)
+    .await?;
+    let total_input: i64 = billing_row.try_get(0).unwrap_or(0);
+    let total_output: i64 = billing_row.try_get(1).unwrap_or(0);
+    let total_requests: i64 = billing_row.try_get(2).unwrap_or(0);
+
+    // === Breakdowns: from usage_records (per-request detail) ===
+    // Optional date filters: pass as TEXT with explicit SQL cast.
     let from_str: Option<&str> = query.from.as_deref();
-    // For 'to' date-only strings, append T23:59:59 to include the entire day
     let to_str: Option<String> = query.to.as_ref().map(|s| {
         if s.len() == 10 { format!("{}T23:59:59", s) } else { s.clone() }
     });
 
-    // Build SQL dynamically to avoid sqlx NULL-without-type-OID problem entirely.
-    // Date parameters are injected as SQL literals (validated above via chrono parse).
-    // Only account_id uses parameterized binding to prevent SQL injection on user input.
+    // Build SQL dynamically for usage_records breakdowns.
+    // Date parameters are injected as SQL literals (validated via chrono parse).
     let mut where_parts = vec![format!("account_id = '{}'", account_id.replace('\'', "''"))];
     if let Some(f) = from_str {
-        // Validate: must be parseable as a date
         let valid = chrono::NaiveDate::parse_from_str(f, "%Y-%m-%d").is_ok()
             || chrono::NaiveDateTime::parse_from_str(f, "%Y-%m-%dT%H:%M:%S%.f").is_ok();
         if !valid {
@@ -457,15 +469,6 @@ pub async fn get_usage_stats(
     }
     let where_clause = where_parts.join(" AND ");
 
-    let total_sql = format!(
-        "SELECT COUNT(*)::bigint, COALESCE(SUM(input_tokens), 0)::bigint, COALESCE(SUM(output_tokens), 0)::bigint
-         FROM usage_records WHERE {}", where_clause
-    );
-    let row = sqlx::query(&total_sql).fetch_one(db).await?;
-    let total_requests: i64 = row.try_get(0).unwrap_or(0);
-    let total_input: i64 = row.try_get(1).unwrap_or(0);
-    let total_output: i64 = row.try_get(2).unwrap_or(0);
-
     // 按模型统计
     let by_model_sql = format!(
         "SELECT provider_id, model_id, COUNT(*)::bigint AS request_count, COALESCE(SUM(input_tokens), 0)::bigint AS input_tokens, COALESCE(SUM(output_tokens), 0)::bigint AS output_tokens