feat: classify stream-failed audits as errors and add picklist filters

[MMortise]

Summary

Fixes a misclassification in request audit: when a request returned a 2xx response header but the stream later failed (e.g. Build response-header timeout / first-byte timeout / upstream_stream_interrupted), the audit record kept status_code = 200. As a result these failed requests were counted as successful in audit summaries, the success filter, and the dashboard, and could not be filtered out.

Backend

• auditStatusForRecord: when 2xx headers were sent but error_code is non-empty, the stored status code becomes 0 (outside any HTTP band). Applied to both the responses path (service.go) and image/video (image.go).
• Audit summary / dashboard: a request is now successful only when status_code is 2xx and error_code is empty; everything else counts as failed.
• New status=other filter (status_code < 100 OR status_code >= 600) so these records can be queried directly.
• Startup migration ensureAuditStatusCodeConstraints: relaxes the chk_request_audits_status_code CHECK to allow 0, then backfills existing 2xx+error_code records to 0. Order matters (constraint first, then backfill). The "allows zero" check is dialect-aware (SQLite BETWEEN 0 AND 599 vs PostgreSQL (status_code >= 0) AND (status_code <= 599)), idempotent on both.

Frontend (request audits page)

• Status column: when statusCode === 0 with an error code, show an amber "error" label instead of the misleading 200; the tooltip keeps the concrete error code.
• Status filter gains an "other error" option.
• Key / account filters upgraded from free-text inputs to searchable, scrollable picklists reusing the existing DataTableFilters group-search pattern:
  • Keys: lists all keys, input matches and click-to-select.
  • Accounts: lists up to the first 50 matches (search re-queries the first 50).
  • The list is capped at ~5 rows and scrolls internally.
  • Accounts query no longer pins provider so accounts from any provider (grok_build / grok_web / grok_console) can be filtered.

Verification

• Backend: 56 test packages pass; added tests for auditStatusForRecord, the other filter + summary semantics, the backfill migration idempotency, and the dialect-aware constraint check.
• Frontend: tsc, eslint, vite build all pass.
• Verified against a copy of a production database: 79 existing 2xx+error_code records are backfilled to 0 and become visible under the "other error" filter.

Screenshot (local acceptance run)

Status column shows the amber error label and the filter picklists for keys/accounts.

[MMortise]

[chenyme]

补充优化

根据代码审查进一步调整审计结果模型与大数据量筛选实现：

• 保留请求真实 HTTP 状态码，不再将“2xx 响应头后流式失败”改写为 status_code = 0
• 使用 status_code + error_code 判断最终结果：
  • 仅 2xx 且无 error_code 计为成功
  • 2xx 后流式中断、首字节超时等计为失败
• “其它错误”筛选可直接查询上述流式失败记录
• 前端以 200 · 错误 展示，兼顾真实 HTTP 状态与最终执行结果
• 移除状态码 CHECK 约束重建和历史全表回填，避免大型审计表在启动时产生锁表、磁盘放大或长时间迁移
• 失败数量由“请求总数－成功数量”计算，减少重复聚合表达式

大账号池优化

• 账号与密钥筛选采用服务端搜索，前端最多保留前 50 条结果，避免无限分页造成 DOM 和内存持续增长
• 用户输入数字 ID 时自动执行内部精确查询，SQLite/PostgreSQL 均可使用主键索引
• 保留普通数字名称的原有模糊搜索语义
• 超出 50 条时提示继续输入关键词缩小范围
• 账号选项包含名称、ID 和 Provider，密钥选项包含名称、ID 和前缀，避免重名歧义

界面优化

• 筛选菜单调整为紧凑的系统级菜单样式
• 使用无边框搜索、轻量分隔和克制的选中态
• 名称保持单行省略，ID 与 Provider 作为次级信息
• 移除重复分组标题和胶囊式 Provider 标签
• 列表保持固定高度并支持内部滚动

验证

• go test ./...
• go vet ./...
• pnpm lint
• pnpm build
• git diff --check
• 5 万条审计记录聚合基准约 40ms，未引入启动迁移或无界内存增长