Reviewed the diff end-to-end (independent sweep + adversarial verification, plus a local npm ci + vitest run of the grader tests — 55 pass). The cross-component contract with the #1395 backend is clean: verdict / root_cause / evidence_type / grader_type serialization, optional-vs-null semantics, the 409 conversation_not_ready shape, /grader/latest returning a bare result or JSON null, and the request body all line up field-for-field. Both new effects converge (no re-fetch loop; the AtifViewer mount effect stays []-gated so widening handleLoad's deps is loop-safe). Nice split.

Two low, non-blocking items:

1. [low] 7 of 12 interruption types render as raw English in the grader panel. findingLabel (EvaluationPanel.tsx:278) and evidenceLabel (:320, via its findingLabel fallback) only translate 5 interruption codes (llm_error, sse_truncated, network_timeout, service_unavailable, agent_crash). But interruption-derived findings carry finding.code = record.interruption_type (grader/rule.rs:277) and their evidence carries label = interruption_type (grader/evidence.rs:155), and the full vocabulary is 12 — all real emitted variants in InterruptionType::as_str (interruption/types.rs:41-51). So findings/evidence for auth_error, context_overflow, dead_loop, rate_limit, retry_storm, safety_filter, token_limit fall through labels[value] ?? value and show raw snake_case, even though apiClient.ts already exports INTERRUPTION_TYPE_CN with Chinese for all of them (and InterruptionBadge already renders those same types translated). Consulting INTERRUPTION_TYPE_CN in the fallback would close the gap.

2. [nit] Duplicate React key in the findings list. key={${finding.code}-${finding.message}} (EvaluationPanel.tsx:160). build_findings emits one finding per interrupted event, and one per unresolved interruption of the same type, so (code, message) is not unique — 2+ such findings produce identical keys and React logs "Encountered two children with the same key". I verified all rows still render on initial mount (no dropped rows), so this is a console-warning / reconciliation-hygiene nit, not a visible bug. A per-item index or a ref id in the key clears it.

Everything else LGTM.

Code Review

整体代码结构清晰，组件拆分合理，测试覆盖充分（95 tests pass）。类型定义完整且与后端 schema 对齐良好。以下是几点建议：

🔴 Finding 1（建议必须修复）— evaluateConversation 未使用 apiFetch 封装

apiClient.ts 中 fetchLatestEvaluation 使用了项目统一的 apiFetch，但 evaluateConversation 直接调用原生 fetch：

const res = await fetch(`${API_BASE}/api/grader/evaluate`, {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ ... }),
});

apiFetch 封装了两个认证关键逻辑：

1. credentials: "same-origin" — 确保携带 session cookie
2. 401 状态码自动跳转登录页

绕过 apiFetch 意味着在 server.auth.enabled = true 的环境下，POST 请求不会携带 cookie，后端返回 401 后用户也不会被引导到登录页，"开始评估"按钮会直接失败。

建议：扩展 apiFetch 支持 POST（如新增 apiFetchPost<T>(url, body)），或让 evaluateConversation 复用相同的 credentials 和 401 处理逻辑。

🟡 Finding 2 — EvaluationPanel 的 initialResult prop 更新后不生效

const [result, setResult] = useState<EvaluationResult | null>(initialResult);

useState(initialValue) 只在首次 mount 时使用初始值，后续 prop 变化不会同步。当前因为面板展开/收起会重新挂载，影响有限，但如果将来需要在面板打开状态下更新数据会出问题。建议添加 useEffect 监听 initialResult 变化。

🟡 Finding 5 — searchParams 加入 useEffect 依赖可能导致不必要的重新请求

AtifViewerPage.tsx 中将 searchParams 加入了 fetch ATIF 文档的 useEffect 依赖数组，但 effect 内部又会调用 setSearchParams，形成潜在的循环触发。建议从依赖中移除 searchParams，改用 useRef 或在回调体内读取最新值。

🟢 其他小建议

• highlightedSections 中 doc.steps 缺少空值防御，建议加 doc.steps ?? []
• Evidence ref 的 React key 可能不唯一，建议加上数组 index 作为 tiebreaker
• EvaluationPanel.tsx 中约 140 行是独立的 label 映射函数，可考虑统一到单一 i18n 对象中（不阻塞合入）

Re-reviewed 6dcfb9a1 — all the raised points are resolved and I verified them against a local build (npm ci, 5 grader test files → 103 passing, tsc --noEmit clean):

• Duplicate finding/evidence keys — now unique (-${index}); confirmed with a discriminating test (two findings sharing code+message no longer emit React's duplicate-key warning).
• Interruption i18n — findingLabel now falls back to INTERRUPTION_TYPE_CN, so the 7 previously-untranslated types (rate_limit, auth_error, …) render in Chinese; evidenceLabel inherits it.
• @chengshuyi's apiFetch point — evaluateConversation now goes through the shared apiFetch, so grader mutations carry credentials: same-origin + the 401→login redirect, and 409 conversation_not_ready maps cleanly via the new structured ApiRequestError. The refactor is backward-compatible for the existing GET/POST/DELETE callers (init defaults to {}, credentials is forced last).
• initialResult sync, the steps ?? [] guards, and the ATIF ref-based URL read are all in and test-covered; the deeplink-highlight path still passes.

Contract with the #1395 backend remains field-aligned. LGTM — approving. @chengshuyi feel free to confirm the apiFetch change reads right on your end.