From 325849816ed86a1faba35e5b8cd55bc80914bcde Mon Sep 17 00:00:00 2001
From: QiuSW <ila2002@qq.com>
Date: Fri, 11 Sep 2026 22:46:23 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20=E6=95=B4=E6=94=B9=20#8=20=E5=AE=A1?=
 =?UTF-8?q?=E6=A0=B8=E9=97=AE=E9=A2=98=20R1=EF=BD=9ER3=20(#8)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- R2 并发同键作答：取得词条行锁后加锁复查答案键，stale 插入遇到唯一键冲突转为返回
  已记录结果，不再返回 500；新增两个 goroutine 同键提交的集成用例
- R3 排期：只有新建或状态/等级实际变化才移动 due_at，编辑释义与例句保留原排期，
  逾期词条不会被挤出当天队列
- R3 附带发现：保存未提及等级时保留已获得的等级，阅读器面板不再把 4 级词重置为 1 级
- R1 契约：作答响应 result 只取 applied/stale，另加 duplicate 标记，重放返回首次结果；
  客户端按首次结果计数，本轮只解决卡片而没有新计分时显示完成页而不是空队列
- R4/R5：stale 与重放分别给出角色为 status 的提示，answerId 作用域注释与实现一致
- Wiki 更新 Business-Rules-and-Glossary、Architecture-and-Code-Map、
  Local-Development-and-Verification 并同步镜像
---
 docs/02-architecture-and-code-map.md          |   6 +-
 docs/03-business-rules-and-glossary.md        |   8 +-
 docs/04-local-development-and-verification.md |  22 ++-
 learner/src/__tests__/review.spec.ts          |  38 +++++-
 learner/src/stores/review.ts                  |  35 ++++-
 learner/src/style.css                         |   1 +
 learner/src/views/ReviewView.vue              |   4 +-
 server/app/lexgo/review.go                    |  76 ++++++++---
 server/app/lexgo/review_test.go               | 125 +++++++++++++++++-
 server/app/lexgo/router.go                    |   2 +-
 server/app/lexgo/terms.go                     |  78 ++++++++---
 server/app/lexgo/terms_test.go                |  55 +++++---
 12 files changed, 362 insertions(+), 88 deletions(-)

diff --git a/docs/02-architecture-and-code-map.md b/docs/02-architecture-and-code-map.md
index 2df66aa..1b293f2 100644
--- a/docs/02-architecture-and-code-map.md
+++ b/docs/02-architecture-and-code-map.md
@@ -2,8 +2,8 @@
 generated: true (请先修改 Gitea Wiki，禁止直接编辑本文件)
 wiki_page: Architecture-and-Code-Map
 wiki_url: https://git.ilapage.cn/OPC/lexgo/wiki/Architecture-and-Code-Map.-
-wiki_revision: 0d6fb43a2662e263fc6f2fb5426b9a8fafaa4868
-synchronized_at: 2026-09-11T12:54:05Z
+wiki_revision: 8add5553222f7f5235c0b3fe45a0ad46d37bbd46
+synchronized_at: 2026-09-11T14:44:50Z
 <!-- gitea-wiki-mirror:end -->
 
 # 架构与代码地图
@@ -267,7 +267,7 @@ schema v6 新增 `lexgo_term_reviews`（每个个人词条一行排期：`due_at
 | 接口 | 权限与输入/输出 |
 |---|---|
 | GET /api/v1/reviews/queue | 本人＋当前语言；仅 `新词`/`学习中` 且 `due_at ≤ now`，按 `due_at, id` 排序、最多 50 条；返回 `{items, total}`，`total` 是全部到期数；不接受查询参数 |
-| POST /api/v1/reviews/:termId/answers | `{answerId, grade: correct\|wrong\|again, expectedDueAt}`；应用成功 201，重放或过期 200；返回 `result`(applied/duplicate/stale)、前后状态/等级/到期时间、`requeued` 与词条新状态 |
+| POST /api/v1/reviews/:termId/answers | `{answerId, grade: correct\|wrong\|again, expectedDueAt}`；应用成功 201，重放或过期 200；返回首次结果 `result`(applied/stale) 与 `duplicate` 标记、前后状态/等级/到期时间、`requeued` 与词条新状态；加锁后再次读取答案键，所以并发的同键提交也返回记录而不是报错 |
 
 学习端新增 `/review` 路由与书库、阅读器顶栏的「到期复习」入口；`stores/review.ts` 维护队列、本轮计数、评分与重学，`ReviewCard.vue`/`ReviewView.vue` 呈现正面（词＋挖空例句）、答案面（个人释义＋例句＋三个评分按钮）、完成页与空队列页。一次评分对应一个 `answerId`，失败重试复用同一个；换词后重新生成。切换账号或退出登录会清空队列、计数与当前卡片。
 
diff --git a/docs/03-business-rules-and-glossary.md b/docs/03-business-rules-and-glossary.md
index 9327b3a..43c3e02 100644
--- a/docs/03-business-rules-and-glossary.md
+++ b/docs/03-business-rules-and-glossary.md
@@ -2,8 +2,8 @@
 generated: true (请先修改 Gitea Wiki，禁止直接编辑本文件)
 wiki_page: Business-Rules-and-Glossary
 wiki_url: https://git.ilapage.cn/OPC/lexgo/wiki/Business-Rules-and-Glossary.-
-wiki_revision: e59184a923910cee9283771154349397e7063fbd
-synchronized_at: 2026-09-11T12:54:06Z
+wiki_revision: 1d90013d883de9339865694e0bb0026f9edd3acc
+synchronized_at: 2026-09-11T14:44:50Z
 <!-- gitea-wiki-mirror:end -->
 
 # 业务规则与术语
@@ -181,13 +181,13 @@ exact优先；未命中再按WordNet异常表/词尾规则查候选，词性顺
 |---|---|---|---|---|---|---|---|
 | 下次复习 | 1 天 | 2 天 | 4 天 | 7 天 | 15 天 | 30 天 | 60 天 |
 
-**入队范围**：只有状态 `新词` 或 `学习中` 且 `due_at ≤ now` 的词条进入队列；`已知` 与 `忽略` 不入队。新保存的词立即到期（`due_at` = 保存时刻），保存后就能复习；手动把词条改为「学习中 level N」时下次复习为 `now + 间隔[N]`，避免刚标等级就被当成到期。已在队列中的词条被改成 `已知`/`忽略` 后再次作答会被拒绝（409），排队信息由服务端裁决而不是前端过滤。
+**入队范围**：只有状态 `新词` 或 `学习中` 且 `due_at ≤ now` 的词条进入队列；`已知` 与 `忽略` 不入队。新保存的词立即到期（`due_at` = 保存时刻），保存后就能复习；手动把词条改为「学习中 level N」时下次复习为 `now + 间隔[N]`，避免刚标等级就被当成到期。只有新建词条或状态/等级实际变化才移动复习时间：只修改释义、例句或原词形时保留原有排期，否则编辑文本会把逾期词条挤出当天队列。保存未提及等级时保留已获得的等级，只有进入 `学习中` 才从 1 开始，因此阅读器面板的保存不会把等级重置为 1。已在队列中的词条被改成 `已知`/`忽略` 后再次作答会被拒绝（409），排队信息由服务端裁决而不是前端过滤。
 
 **三个评分动作**：`correct` 认识/答对 → 等级 +1（封顶 7）、按新等级排期、离开本轮；`wrong` 不认识/答错 → `学习中` 降一级（最低 1）、`新词` 保持 `新词`、`due_at = now` 立即回到本轮；`again` 再学一次 → 等级与状态不变、`due_at = now` 立即回到本轮。计数上 `correct_count` 只统计 `correct`，`wrong_count` 统计 `wrong` 与 `again`，`review_count` 统计全部已应用作答。词条的原文、个人释义与例句不因复习改变。
 
 **时区与到期边界**：`due_at` 以 UTC 绝对时刻存储，到期判定是 `due_at ≤ now`，不引入本地日边界。理由：MVP 没有用户时区设置（属 X 系列边界），绝对时刻在多账号自托管下语义一致、没有夏令时陷阱；代价是复习时刻会随首次作答时间漂移，例如 22:00 答对的 1 天间隔词条要到次日 22:00 才到期。原型界面的「到期复习 N / M」指当前到期队列的位置与本轮卡片数，不是自然日统计。
 
-**幂等与并发**：客户端每次作答生成一个 `answerId`，服务端以 `UNIQUE(owner_id, answer_id)` 去重。同一个 `answerId` 再次提交返回首次结果（`result=duplicate`）且不改变等级、间隔和次数；同一词条在别处已经被推进（提交回传的 `expectedDueAt` 与服务端当前 `due_at` 不一致）时记为 `result=stale`，同样不改变任何状态。因此网络重发、双击、双标签页作答都只记账一次。每次尝试都会落一条 `lexgo_review_answers`（含 `result`），这是「只记账一次」的证据，也供 #13 统计使用。
+**幂等与并发**：客户端每次作答生成一个 `answerId`，服务端以 `UNIQUE(owner_id, answer_id)` 去重。同一个 `answerId` 再次提交返回首次结果并把 `duplicate` 标记为 true（`result` 仍是首次的 `applied` 或 `stale`），不改变等级、间隔和次数，因此客户端重试可以按首次结果计数；同一词条在别处已经被推进（提交回传的 `expectedDueAt` 与服务端当前 `due_at` 不一致）时记为 `result=stale`，同样不改变任何状态。因此网络重发、双击、双标签页作答都只记账一次。每次尝试都会落一条 `lexgo_review_answers`（含 `result`），这是「只记账一次」的证据，也供 #13 统计使用。
 
 **归属与错误**：词条归属由服务端按会话裁决，`owner` 不接受客户端输入；他人词条与不存在的词条统一 404，未登录 401，未知评分/缺少 `expectedDueAt`/未知字段 400，词条已变成 `已知`/`忽略` 409。复习接口不写审计日志：答题属于私人学习内容。
 
diff --git a/docs/04-local-development-and-verification.md b/docs/04-local-development-and-verification.md
index 8778409..dacbc51 100644
--- a/docs/04-local-development-and-verification.md
+++ b/docs/04-local-development-and-verification.md
@@ -2,8 +2,8 @@
 generated: true (请先修改 Gitea Wiki，禁止直接编辑本文件)
 wiki_page: Local-Development-and-Verification
 wiki_url: https://git.ilapage.cn/OPC/lexgo/wiki/Local-Development-and-Verification.-
-wiki_revision: e01464ad570bbd037e813985d78b137462bca241
-synchronized_at: 2026-09-11T12:54:06Z
+wiki_revision: 1192d23a8198e981961adfc6066f0b3bc3b84058
+synchronized_at: 2026-09-11T14:44:50Z
 <!-- gitea-wiki-mirror:end -->
 
 # 本地开发与验证
@@ -362,3 +362,21 @@ node --test spikes/english/view.test.mjs
 运维记录：本次只有 lexgo-api（新二进制）与 lexgo-learner 被重启；学习端常驻 Vite 在长时间运行后一度对 `/src/style.css` 返回空样式表，导致 E2E 观察到 `white-space: normal`，重启后恢复，未改动代码或配置。其余实例保持 Running。
 
 未验证：真实手机触屏详细证据与完整备份恢复演练仍属既有缺口（#14/#15）；本单只用桌面浏览器窄屏检查，不当作真机结果。并发只覆盖到「双标签页同一词条」这一层，没有做多用户压测。
+
+## #8 审核整改（R1～R3，2026-09-11）
+
+独立审核（Claude Code）只读审阅提交 `0328505`／`ec5ec2d`，指出三处影响验收标准第 2、3 条的问题。三点均先在 `lexgo_test_issue8` 写复现用例观察到失败，再修复并转绿。
+
+| 问题 | 现象与根因 | 修复 | 回归用例 |
+|---|---|---|---|
+| R2 并发同键返回 500 | 两个同时到达、带同一 `answerId` 的请求都在加锁前查不到记录；后者进入 stale 分支插入相同 `answer_key`，触发唯一键冲突返回 500 | 取得词条行锁后再用加锁读复查一次答案键，命中直接返回首次结果；stale 插入遇到 1062 也转为返回记录 | `TestMySQLReviewConcurrentReplayOfOneAnswer`（两个 goroutine 同键提交，两个都 2xx、`review_count = 1`、只有一条答案记录） |
+| R3 编辑文本会重排复习 | `saveTerm` 无条件调用 `syncTermReview`，编辑释义/例句也会把 `due_at` 重算，逾期词条被挤出当天队列 | 保存前加锁读取旧行，只有新建或状态/等级实际变化才移动 `due_at`；缺行时补建排期行 | `TestMySQLReviewEditKeepsSchedule`（逾期 3 级词只改释义：`due_at` 与队列不变；改等级则重排） |
+| R3 附带发现：面板保存把等级重置为 1 | 阅读器面板只提交状态不提交等级，`termLevel` 对缺省等级一律返回 1，于是 4 级词改一个错字会掉到 1 级 | 保存未提及等级时保留已获得的等级；只有进入 `学习中` 才从 1 开始 | `TestMySQLReviewPanelSaveKeepsLevel`（4 级词面板式保存后仍为 4 级，且排期不变；退出再进入学习中则从 1 开始） |
+
+**答案契约随之明确**：作答响应 `result` 只取 `applied`/`stale`，另加 `duplicate` 布尔标记。重放返回首次结果并把 `duplicate` 置真，客户端因此可以按首次结果计数：网络把响应丢掉后点「重试提交」拿到 `duplicate=true` 的 `applied`，本轮计数正常增加，完成页不会退化成「今天没有到期词条」。已应用的作答返回 201，重放与 stale 返回 200（R1）。
+
+**提示与注释（R4、R5）**：卡片因 `stale` 离开时页面显示 `role="status"` 提示「该词已在其他页面复习，本次未计分。」，重放且首次为 stale 时显示「该词已按上一次的评分记录，未重复计分。」；本轮只解决卡片而没有新计分时，完成页显示「本轮没有新的计分：N 个词条已在其他页面复习。」。`answerId` 的作用域注释改为「每张卡片一个，失败重试复用」，与 `answerIdFor` 的实现一致。
+
+**流程记录（R6）**：评论 7769 的方案写的是在 `lexgo_terms` 上增加列，实际实现改为独立表 `lexgo_term_reviews`（加法迁移可重试、不对既有表做 ALTER）。该变更在实施评论 7776 与 Wiki 中说明了原因，但没有按「数据结构变化先更新工单」的要求在实施前追加变更评论；本页与上文契约按实际实现记录，方案评论中的「新增列均有默认值」以独立表为准。
+
+整改后重跑：Go 单元与集成测试（专用库 `lexgo_test_issue8`）44 个顶层用例全部通过、0 跳过；学习端 73 项单测、类型检查、构建与 5 项 E2E 通过；管理端 31 项与 lint 通过；治理 56 项与严格检查通过；真实 API＋MySQL 42 项检查通过（新增 4 项针对 R3 与重放契约）；真实浏览器复核面板保存与复习闭环通过。截图 `.local/evidence/issue8-fixed-summary.png`。
diff --git a/learner/src/__tests__/review.spec.ts b/learner/src/__tests__/review.spec.ts
index 3d6bc63..3d0edd6 100644
--- a/learner/src/__tests__/review.spec.ts
+++ b/learner/src/__tests__/review.spec.ts
@@ -69,7 +69,7 @@ describe('review store', () => {
     expect(review.finished).toBe(false)
   })
 
-  it('treats a stale or replayed answer as the same action, never as a second review', async () => {
+  it('treats a stale answer as no new score but finishes the round instead of reporting an empty queue', async () => {
     vi.spyOn(globalThis, 'fetch').mockImplementation(async input => String(input).endsWith('/reviews/queue') ? ok({ items: [item()], total: 1 }) : ok(answer({ result: 'stale', requeued: false })))
     const review = useReviewStore()
     await review.load()
@@ -77,7 +77,33 @@ describe('review store', () => {
     expect(review.queue).toHaveLength(0)
     expect(review.answered).toBe(0)
     expect(review.correctCount).toBe(0)
-    expect(review.empty).toBe(true)
+    expect(review.resolved).toBe(1)
+    expect(review.empty).toBe(false)
+    expect(review.finished).toBe(true)
+    expect(review.notice).toContain('已在其他页面复习')
+  })
+
+  it('counts a replay of this client own answer after a lost response', async () => {
+    let sent = 0
+    vi.spyOn(globalThis, 'fetch').mockImplementation(async input => {
+      if (String(input).endsWith('/reviews/queue')) return ok({ items: [item()], total: 1 })
+      sent += 1
+      // The first response never reaches the client, the retry reports the recorded answer.
+      if (sent === 1) throw new Error('网络中断')
+      return ok(answer({ result: 'applied', duplicate: true }))
+    })
+    const review = useReviewStore()
+    await review.load()
+    await review.answer('correct')
+    expect(review.error).toContain('网络中断')
+    expect(review.answered).toBe(0)
+    await review.answer('correct')
+    expect(review.answered).toBe(1)
+    expect(review.correctCount).toBe(1)
+    expect(review.wordsReviewed).toBe(1)
+    expect(review.empty).toBe(false)
+    expect(review.finished).toBe(true)
+    expect(review.notice).toBe('')
   })
 
   it('keeps the card and the same answer id when a submission fails, then retries once', async () => {
@@ -203,6 +229,14 @@ describe('review page', () => {
     expect(view.text()).toContain('还有 2 个词条到期')
   })
 
+  it('tells the learner when a card was already reviewed elsewhere', async () => {
+    const { view } = await open({ items: [item()], total: 1 }, () => ok(answer({ result: 'stale', duplicate: false })))
+    await view.get('[data-testid="review-reveal"]').trigger('click'); await flushPromises()
+    await view.get('[data-testid="review-correct"]').trigger('click'); await flushPromises()
+    expect(view.get('[data-testid="review-notice"]').text()).toContain('已在其他页面复习')
+    expect(view.get('[data-testid="review-summary"]').text()).toContain('本轮没有新的计分')
+  })
+
   it('ends the round without submitting anything', async () => {
     const { view, router, fetchMock } = await open({ items: [item()], total: 1 })
     await view.get('[data-testid="review-end"]').trigger('click'); await flushPromises()
diff --git a/learner/src/stores/review.ts b/learner/src/stores/review.ts
index 189ea58..7b7ea59 100644
--- a/learner/src/stores/review.ts
+++ b/learner/src/stores/review.ts
@@ -3,7 +3,9 @@ import { defineStore } from 'pinia'
 import { useSessionStore } from './session'
 
 export type ReviewGrade = 'correct' | 'wrong' | 'again'
-export type ReviewResult = 'applied' | 'duplicate' | 'stale'
+// The outcome of an answer: applied when the word moved, stale when another screen had
+// already reviewed it. A replayed answer repeats the outcome it was given first.
+export type ReviewResult = 'applied' | 'stale'
 
 export interface ReviewItem {
   id: number
@@ -19,6 +21,7 @@ export interface ReviewItem {
 
 export interface ReviewAnswerResult {
   result: ReviewResult
+  duplicate: boolean
   grade: ReviewGrade
   requeued: boolean
   statusBefore: string
@@ -67,19 +70,25 @@ export const useReviewStore = defineStore('review', () => {
   const wrongCount = ref(0)
   // Distinct words in this round: a requeued word is answered again but is one word.
   const wordsReviewed = ref(0)
+  // Cards this round took off the queue, counted for every outcome, and the reason a card
+  // left without a new score.
+  const resolved = ref(0)
+  const notice = ref('')
   const seen = new Set<number>()
   // Words due beyond the fetched page, reported by the server for this round.
   const pending = ref(0)
   let started = ref(false)
-  // One answer id per card and grade: a retry of the same action keeps its key, so the
-  // server can answer it from the first outcome instead of counting twice.
+  // One answer id per card: a retry of a failed submission reuses its key, so the server
+  // answers the retry from the first outcome instead of counting the same action twice.
   let attempt: { itemId: number; answerId: string } | null = null
   let sequence = 0
   let generation = 0
 
   const current = computed<ReviewItem | null>(() => queue.value[0] ?? null)
-  const finished = computed(() => started.value && !loading.value && queue.value.length === 0 && answered.value > 0)
-  const empty = computed(() => started.value && !loading.value && queue.value.length === 0 && answered.value === 0)
+  // A round that resolved cards is finished even when every answer turned out to be a
+  // replay or a stale submission; only a round that never had a card is empty.
+  const finished = computed(() => started.value && !loading.value && queue.value.length === 0 && resolved.value > 0)
+  const empty = computed(() => started.value && !loading.value && queue.value.length === 0 && resolved.value === 0)
 
   watch(() => session.user?.id ?? null, (next, previous) => {
     if (next !== previous) reset()
@@ -105,6 +114,8 @@ export const useReviewStore = defineStore('review', () => {
     correctCount.value = 0
     wrongCount.value = 0
     wordsReviewed.value = 0
+    resolved.value = 0
+    notice.value = ''
     pending.value = 0
     started.value = false
     attempt = null
@@ -118,6 +129,7 @@ export const useReviewStore = defineStore('review', () => {
     loading.value = true
     error.value = ''
     revealed.value = false
+    notice.value = ''
     attempt = null
     try {
       const result = await session.request<QueueResponse>('reviews/queue')
@@ -133,6 +145,7 @@ export const useReviewStore = defineStore('review', () => {
         correctCount.value = 0
         wrongCount.value = 0
         wordsReviewed.value = 0
+        resolved.value = 0
         seen.clear()
       }
       started.value = true
@@ -185,7 +198,11 @@ export const useReviewStore = defineStore('review', () => {
     queue.value = queue.value.filter(entry => entry.id !== item.id)
     attempt = null
     revealed.value = false
+    resolved.value += 1
     if (result.result === 'applied') {
+      // A replay repeats the first outcome, and the first attempt may be this client's own
+      // submission whose response was lost, so it counts exactly like that attempt.
+      notice.value = ''
       answered.value += 1
       if (!seen.has(item.id)) {
         seen.add(item.id)
@@ -193,6 +210,10 @@ export const useReviewStore = defineStore('review', () => {
       }
       if (result.grade === 'correct') correctCount.value += 1
       else wrongCount.value += 1
+    } else {
+      notice.value = result.duplicate
+        ? '该词已按上一次的评分记录，未重复计分。'
+        : '该词已在其他页面复习，本次未计分。'
     }
     if (result.requeued) {
       const next = result.item ?? { ...item, level: result.levelAfter, status: result.statusAfter as ReviewItem['status'], dueAt: result.dueAtAfter }
@@ -206,7 +227,7 @@ export const useReviewStore = defineStore('review', () => {
   }
 
   return {
-    queue, current, loading, error, busy, revealed, answered, correctCount, wrongCount, wordsReviewed, pending,
-    finished, empty, load, reveal, answer, continueRound, reset,
+    queue, current, loading, error, busy, revealed, answered, correctCount, wrongCount, wordsReviewed, resolved,
+    notice, pending, finished, empty, load, reveal, answer, continueRound, reset,
   }
 })
diff --git a/learner/src/style.css b/learner/src/style.css
index 04a7eba..d7fe386 100644
--- a/learner/src/style.css
+++ b/learner/src/style.css
@@ -128,6 +128,7 @@ a.chapter-name:hover { color: #315c43; text-decoration: underline; }
 .lookup-saved { color: #2f6b45; font-size: 14px; margin: 12px 0 0; }
 .lookup-actions { display: flex; gap: 8px; margin-top: 14px; flex-wrap: wrap; }
 .review-page { max-width: 680px; }
+.review-notice { margin: 10px 0 0; padding: 10px 14px; border: 1px solid #d9decf; border-radius: 8px; background: #fbf7ee; color: #6b5b3e; }
 .review-card, .review-summary { margin-top: 26px; padding: 28px; border: 1px solid #d9decf; border-radius: 14px; background: #fffdf8; }
 .review-summary h2 { margin-top: 0; font-family: Georgia, serif; font-size: 24px; }
 .review-word { margin: 12px 0; font-family: Georgia, serif; font-size: 34px; }
diff --git a/learner/src/views/ReviewView.vue b/learner/src/views/ReviewView.vue
index dd6e3b1..7d55e99 100644
--- a/learner/src/views/ReviewView.vue
+++ b/learner/src/views/ReviewView.vue
@@ -38,6 +38,7 @@ const requireItem = (value: ReviewItem | null): ReviewItem => value as ReviewIte
     </header>
     <main class="page review-page">
       <h1>到期复习</h1>
+      <p v-if="review.notice" role="status" class="review-notice" data-testid="review-notice">{{ review.notice }}</p>
       <p v-if="review.loading" role="status" class="loading">正在加载…</p>
       <div v-else-if="review.error && !review.current" class="notice">
         <p role="alert">{{ review.error }}</p>
@@ -45,7 +46,8 @@ const requireItem = (value: ReviewItem | null): ReviewItem => value as ReviewIte
       </div>
       <section v-else-if="review.finished" class="review-summary" data-testid="review-summary">
         <h2>本次复习完成</h2>
-        <p>复习了 {{ review.wordsReviewed }} 个词条 · 共 {{ review.answered }} 次作答</p>
+        <p v-if="review.answered">复习了 {{ review.wordsReviewed }} 个词条 · 共 {{ review.answered }} 次作答</p>
+        <p v-else class="subtle">本轮没有新的计分：{{ review.resolved }} 个词条已在其他页面复习。</p>
         <p class="subtle">答对 {{ review.correctCount }} · 答错或再学 {{ review.wrongCount }} · 已更新复习计划</p>
         <p v-if="review.pending" role="status" class="notice">还有 {{ review.pending }} 个词条到期。</p>
         <div class="review-grades">
diff --git a/server/app/lexgo/review.go b/server/app/lexgo/review.go
index 0cb34d4..3253749 100644
--- a/server/app/lexgo/review.go
+++ b/server/app/lexgo/review.go
@@ -5,6 +5,7 @@ import (
 	"strconv"
 	"time"
 
+	driver "github.com/go-sql-driver/mysql"
 	"github.com/gin-gonic/gin"
 	admin "go-admin/app/admin/models"
 	"gorm.io/gorm"
@@ -95,7 +96,11 @@ type ReviewAnswerInput struct {
 }
 
 type ReviewAnswerResult struct {
-	Result       string     `json:"result"`
+	// Result is the outcome of this answer: applied when the term moved, stale when another
+	// screen already advanced it. A replayed answer repeats the outcome it was given first.
+	Result string `json:"result"`
+	// Duplicate reports that this answer key was already recorded and nothing changed now.
+	Duplicate    bool       `json:"duplicate"`
 	Grade        string     `json:"grade"`
 	Requeued     bool       `json:"requeued"`
 	StatusBefore string     `json:"statusBefore"`
@@ -206,18 +211,25 @@ func ReviewQueueFor(tx *gorm.DB, owner int, language string, now time.Time) (Rev
 
 // syncTermReview keeps the scheduling row in step with a saved term. A new word is due
 // immediately so saving it starts the loop; a word saved straight into a level waits for
-// that level's interval, so a manual level is not silently re-queued today.
-func syncTermReview(tx *gorm.DB, term Term, now time.Time) error {
+// that level's interval, so a manual level is not silently re-queued today. When the caller
+// does not ask for a reschedule, only a missing row is created and the existing date and
+// counters stay untouched.
+func syncTermReview(tx *gorm.DB, term Term, reschedule bool, now time.Time) error {
 	due := stamp(now)
 	if term.Status == termStatusLearning && term.Level >= 1 {
 		due = nextDue(term.Level, now)
 	}
-	row := TermReview{TermID: term.ID, OwnerID: term.OwnerID, Language: term.Language, DueAt: due}
-	// Counters are history and survive a status or level change.
-	return tx.Clauses(clause.OnConflict{
-		Columns:   []clause.Column{{Name: "term_id"}},
-		DoUpdates: clause.Assignments(map[string]any{"due_at": row.DueAt}),
-	}).Create(&row).Error
+	var existing TermReview
+	err := tx.Clauses(clause.Locking{Strength: "UPDATE"}).Where("term_id = ?", term.ID).First(&existing).Error
+	switch {
+	case errors.Is(err, gorm.ErrRecordNotFound):
+		return tx.Create(&TermReview{TermID: term.ID, OwnerID: term.OwnerID, Language: term.Language, DueAt: due}).Error
+	case err != nil:
+		return err
+	case reschedule:
+		return tx.Model(&TermReview{}).Where("term_id = ?", term.ID).Update("due_at", due).Error
+	}
+	return nil
 }
 
 func lockedReview(tx *gorm.DB, owner int, term Term) (TermReview, error) {
@@ -234,6 +246,15 @@ func lockedReview(tx *gorm.DB, owner int, term Term) (TermReview, error) {
 	return review, err
 }
 
+// lockedAnswerKey reads one recorded answer under a lock. It is called again after the term
+// row is locked, because a locking read sees the newest committed row while the plain read
+// before it may still see this transaction's snapshot.
+func lockedAnswerKey(tx *gorm.DB, owner int, key string) (ReviewAnswer, error) {
+	var stored ReviewAnswer
+	err := tx.Clauses(clause.Locking{Strength: "UPDATE"}).Where("owner_id = ? AND answer_key = ?", owner, key).First(&stored).Error
+	return stored, err
+}
+
 // AnswerReview applies one grade to one owned term exactly once. A repeated answer key
 // returns the first outcome, and an answer whose due time no longer matches (another tab
 // already advanced the term) is recorded as stale without moving anything.
@@ -248,15 +269,9 @@ func AnswerReview(tx *gorm.DB, owner int, termID int64, input ReviewAnswerInput,
 	if err != nil {
 		return ReviewAnswerResult{}, err
 	}
-	var stored ReviewAnswer
-	err = tx.Where("owner_id = ? AND answer_key = ?", owner, key).First(&stored).Error
-	if err == nil {
-		if stored.TermID != termID {
-			return ReviewAnswerResult{}, failure(400, "请求编号已用于其他词条")
-		}
-		return replayedAnswer(tx, stored)
-	}
-	if !errors.Is(err, gorm.ErrRecordNotFound) {
+	if stored, err := lockedAnswerKey(tx, owner, key); err == nil {
+		return replayedAnswer(tx, stored, termID)
+	} else if !errors.Is(err, gorm.ErrRecordNotFound) {
 		return ReviewAnswerResult{}, err
 	}
 	// Ownership first: another account's term and a missing term answer identically.
@@ -266,6 +281,13 @@ func AnswerReview(tx *gorm.DB, owner int, termID int64, input ReviewAnswerInput,
 	} else if err != nil {
 		return ReviewAnswerResult{}, err
 	}
+	// Simultaneous submissions of one answer id are serialized by the term lock above; the
+	// second one now sees the recorded answer and reports it instead of counting again.
+	if stored, err := lockedAnswerKey(tx, owner, key); err == nil {
+		return replayedAnswer(tx, stored, termID)
+	} else if !errors.Is(err, gorm.ErrRecordNotFound) {
+		return ReviewAnswerResult{}, err
+	}
 	review, err := lockedReview(tx, owner, term)
 	if err != nil {
 		return ReviewAnswerResult{}, err
@@ -315,7 +337,6 @@ func AnswerReview(tx *gorm.DB, owner int, termID int64, input ReviewAnswerInput,
 		DueAtBefore: before.DueAt, DueAtAfter: next.DueAt, Item: reviewItem(term, review),
 	}, nil
 }
-
 // staleAnswer records that this attempt changed nothing because the term had already
 // moved on. It is a normal outcome of two open tabs, not an error.
 func staleAnswer(tx *gorm.DB, owner int, term Term, review TermReview, input ReviewAnswerInput, now time.Time) (ReviewAnswerResult, error) {
@@ -329,6 +350,14 @@ func staleAnswer(tx *gorm.DB, owner int, term Term, review TermReview, input Rev
 		DueAtBefore: review.DueAt, DueAtAfter: review.DueAt, CreatedAt: stamp(now),
 	}
 	if err := tx.Create(&record).Error; err != nil {
+		// Losing a race to another transaction that recorded this very answer id is not an
+		// error: report the outcome that was stored first.
+		var duplicate *driver.MySQLError
+		if errors.As(err, &duplicate) && duplicate.Number == 1062 {
+			if stored, readErr := lockedAnswerKey(tx, owner, key); readErr == nil {
+				return replayedAnswer(tx, stored, term.ID)
+			}
+		}
 		return ReviewAnswerResult{}, err
 	}
 	return ReviewAnswerResult{
@@ -338,9 +367,12 @@ func staleAnswer(tx *gorm.DB, owner int, term Term, review TermReview, input Rev
 	}, nil
 }
 
-// replayedAnswer returns the outcome already recorded for this answer key. Nothing is
+// replayedAnswer reports the outcome already recorded for this answer key. Nothing is
 // advanced a second time, so a resend cannot change counts or intervals.
-func replayedAnswer(tx *gorm.DB, stored ReviewAnswer) (ReviewAnswerResult, error) {
+func replayedAnswer(tx *gorm.DB, stored ReviewAnswer, termID int64) (ReviewAnswerResult, error) {
+	if stored.TermID != termID {
+		return ReviewAnswerResult{}, failure(400, "请求编号已用于其他词条")
+	}
 	var term Term
 	if err := tx.Where("id = ? AND owner_id = ?", stored.TermID, stored.OwnerID).First(&term).Error; err != nil {
 		return ReviewAnswerResult{}, err
@@ -350,7 +382,7 @@ func replayedAnswer(tx *gorm.DB, stored ReviewAnswer) (ReviewAnswerResult, error
 		return ReviewAnswerResult{}, err
 	}
 	return ReviewAnswerResult{
-		Result: "duplicate", Grade: stored.Grade, Requeued: stored.Requeued,
+		Result: stored.Result, Duplicate: true, Grade: stored.Grade, Requeued: stored.Requeued,
 		StatusBefore: stored.StatusBefore, StatusAfter: stored.StatusAfter,
 		LevelBefore: stored.LevelBefore, LevelAfter: stored.LevelAfter,
 		DueAtBefore: stored.DueAtBefore, DueAtAfter: stored.DueAtAfter, Item: reviewItem(term, review),
diff --git a/server/app/lexgo/review_test.go b/server/app/lexgo/review_test.go
index 5c7776b..2bf1511 100644
--- a/server/app/lexgo/review_test.go
+++ b/server/app/lexgo/review_test.go
@@ -1,6 +1,7 @@
 package lexgo
 
 import (
+	"sync"
 	"encoding/json"
 	"fmt"
 	"testing"
@@ -272,16 +273,17 @@ func TestMySQLReviewAnswerIdempotencyAndStaleTabs(t *testing.T) {
 		t.Fatalf("first tab: %d %#v", code, applied)
 	}
 	code, stale := answerReview(t, r, token, termID, answerBody("tab-two-answer-0002", reviewGradeCorrect, seen))
-	if code != 200 || stale.Result != "stale" || stale.Requeued {
+	if code != 200 || stale.Result != "stale" || stale.Duplicate || stale.Requeued {
 		t.Fatalf("second tab must not advance the term: %d %#v", code, stale)
 	}
 	if stale.LevelAfter != applied.LevelAfter || !stale.DueAtAfter.Equal(applied.DueAtAfter) {
 		t.Fatalf("a stale answer changed state: %#v", stale)
 	}
 
-	// A resend of the first answer returns the first outcome without another update.
+	// A resend of the first answer returns the first outcome, flagged as a replay, without
+	// another update. The client can therefore count it exactly like the original answer.
 	code, duplicate := answerReview(t, r, token, termID, answerBody("tab-one-answer-0001", reviewGradeCorrect, seen))
-	if code != 200 || duplicate.Result != "duplicate" {
+	if code != 200 || duplicate.Result != "applied" || !duplicate.Duplicate {
 		t.Fatalf("resend: %d %#v", code, duplicate)
 	}
 	if duplicate.LevelAfter != applied.LevelAfter || !duplicate.DueAtAfter.Equal(applied.DueAtAfter) || duplicate.Item.ReviewCount != 1 {
@@ -380,3 +382,120 @@ func TestMySQLReviewScheduleOnManualStatus(t *testing.T) {
 		t.Fatalf("a new word is due at once: %#v", queue)
 	}
 }
+
+// TestMySQLReviewConcurrentReplayOfOneAnswer reproduces the review finding that two
+// simultaneous submissions of one answer id must both get the recorded outcome.
+func TestMySQLReviewConcurrentReplayOfOneAnswer(t *testing.T) {
+	db := testDB(t)
+	r, clock, _, token, chapter := reviewFixture(t, db, "Dogs went home.")
+	dogs := saveWord(t, r, token, chapter.ID, 0, 4, termStatusNew, nil)
+	_, queue := reviewQueue(t, r, token)
+	seen := queue.Items[0].DueAt
+	body := answerBody("concurrent-answer-01", reviewGradeCorrect, seen)
+
+	var wg sync.WaitGroup
+	codes := make(chan int, 2)
+	results := make(chan string, 2)
+	for i := 0; i < 2; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			code, data := callAPI(t, r, "POST", fmt.Sprintf("/api/v1/reviews/%d/answers", dogs.Term.ID), token, body)
+			var result ReviewAnswerResult
+			if data != nil {
+				json.Unmarshal(data, &result)
+			}
+			codes <- code
+			results <- result.Result
+		}()
+	}
+	wg.Wait()
+	close(codes)
+	close(results)
+	ok, bad := 0, 0
+	for code := range codes {
+		if code == 200 || code == 201 {
+			ok++
+		} else {
+			bad++
+			t.Logf("status %d", code)
+		}
+	}
+	kinds := []string{}
+	for kind := range results {
+		kinds = append(kinds, kind)
+	}
+	if ok != 2 || bad != 0 {
+		t.Fatalf("both submissions must succeed: ok=%d bad=%d kinds=%v clock=%v", ok, bad, kinds, *clock)
+	}
+	review := termReviewRow(t, db, dogs.Term.ID)
+	if review.ReviewCount != 1 {
+		t.Fatalf("concurrent replay counted %d times", review.ReviewCount)
+	}
+	var attempts int64
+	if err := db.Model(&ReviewAnswer{}).Where("term_id = ?", dogs.Term.ID).Count(&attempts).Error; err != nil || attempts != 1 {
+		t.Fatalf("concurrent replay logged %d attempts", attempts)
+	}
+}
+
+// TestMySQLReviewEditKeepsSchedule reproduces the review finding that editing only the
+// learner's own text must not move a review date (decision D2).
+func TestMySQLReviewEditKeepsSchedule(t *testing.T) {
+	db := testDB(t)
+	r, clock, _, token, chapter := reviewFixture(t, db, "Dogs went home.")
+	level := 3
+	learning := saveWord(t, r, token, chapter.ID, 0, 4, termStatusLearning, &level)
+	// Make the word overdue, as it would be after a missed day.
+	overdue := clock.AddDate(0, 0, -5)
+	if err := db.Model(&TermReview{}).Where("term_id = ?", learning.Term.ID).Update("due_at", overdue).Error; err != nil {
+		t.Fatal(err)
+	}
+	// Editing only the personal text must keep the schedule exactly as it was.
+	code, saved := saveTermAPI(t, r, token, map[string]any{"chapterId": chapter.ID, "start": 0, "end": 4, "definition": "改过的释义", "status": termStatusLearning, "level": 3})
+	if code != 200 || saved.Term.Level != 3 {
+		t.Fatalf("text edit: %d %#v", code, saved)
+	}
+	if review := termReviewRow(t, db, learning.Term.ID); !review.DueAt.Equal(overdue) {
+		t.Fatalf("editing the text moved the due time: %v want %v", review.DueAt, overdue)
+	}
+	if _, queue := reviewQueue(t, r, token); queue.Total != 1 {
+		t.Fatalf("an overdue word must stay due after a text edit: %#v", queue)
+	}
+	// Changing the level is a scheduling action and does move the due time.
+	if code, _ := saveTermAPI(t, r, token, map[string]any{"chapterId": chapter.ID, "start": 0, "end": 4, "status": termStatusLearning, "level": 5}); code != 200 {
+		t.Fatal("level change failed")
+	}
+	if review := termReviewRow(t, db, learning.Term.ID); !review.DueAt.Equal(clock.AddDate(0, 0, 15)) {
+		t.Fatalf("a level change must reschedule: %v", review.DueAt)
+	}
+}
+
+// TestMySQLReviewPanelSaveKeepsLevel reproduces the related defect found while checking the
+// review: the reader panel saves a status without a level, and that must not reset a level
+// the learner already earned.
+func TestMySQLReviewPanelSaveKeepsLevel(t *testing.T) {
+	db := testDB(t)
+	r, clock, _, token, chapter := reviewFixture(t, db, "Dogs went home.")
+	level := 4
+	learning := saveWord(t, r, token, chapter.ID, 0, 4, termStatusLearning, &level)
+	overdue := clock.AddDate(0, 0, -2)
+	if err := db.Model(&TermReview{}).Where("term_id = ?", learning.Term.ID).Update("due_at", overdue).Error; err != nil {
+		t.Fatal(err)
+	}
+	// This is exactly what the reader panel sends: status only, no level.
+	code, saved := saveTermAPI(t, r, token, map[string]any{"chapterId": chapter.ID, "start": 0, "end": 4, "definition": "改过的释义", "status": termStatusLearning})
+	if code != 200 || saved.Term.Level != 4 {
+		t.Fatalf("a panel save must keep the learned level: %d %#v", code, saved)
+	}
+	if review := termReviewRow(t, db, learning.Term.ID); !review.DueAt.Equal(overdue) {
+		t.Fatalf("a panel save moved the due time: %v want %v", review.DueAt, overdue)
+	}
+	// Leaving learning and re-entering it is a real transition and starts at level 1.
+	if code, _ := saveTermAPI(t, r, token, map[string]any{"chapterId": chapter.ID, "start": 0, "end": 4, "status": termStatusNew}); code != 200 {
+		t.Fatal("reset to new failed")
+	}
+	code, again := saveTermAPI(t, r, token, map[string]any{"chapterId": chapter.ID, "start": 0, "end": 4, "status": termStatusLearning})
+	if code != 200 || again.Term.Level != 1 {
+		t.Fatalf("re-entering learning starts at 1: %d %#v", code, again)
+	}
+}
diff --git a/server/app/lexgo/router.go b/server/app/lexgo/router.go
index e9b515f..636fc64 100644
--- a/server/app/lexgo/router.go
+++ b/server/app/lexgo/router.go
@@ -143,7 +143,7 @@ func Router(db *gorm.DB, now func() time.Time) *gin.Engine {
 			}
 			// A replayed or stale answer changed nothing, so it is not a new answer.
 			if c.Request.Method == "POST" && c.FullPath() == "/api/v1/reviews/:termId/answers" {
-				if answer, ok := data.(ReviewAnswerResult); ok && answer.Result == "applied" {
+				if answer, ok := data.(ReviewAnswerResult); ok && answer.Result == "applied" && !answer.Duplicate {
 					status = 201
 				}
 			}
diff --git a/server/app/lexgo/terms.go b/server/app/lexgo/terms.go
index 1d3bc91..45d7be3 100644
--- a/server/app/lexgo/terms.go
+++ b/server/app/lexgo/terms.go
@@ -100,21 +100,25 @@ func termView(t Term) TermView {
 	return TermView{t.ID, t.Language, t.Term, t.OriginalForm, t.Definition, splitExamples(t.Examples), t.Status, t.Level, t.CreatedAt, t.UpdatedAt}
 }
 
-// termLevel enforces the documented status/level boundary: only a learning entry
-// carries a level, entering learning defaults to 1, and every other status must
-// leave the level at 0.
-func termLevel(status string, level *int) (int, error) {
+// termLevel enforces the documented status/level boundary: only a learning entry carries a
+// level, every other status must leave the level at 0, and entering learning starts at 1.
+// A save that does not mention a level keeps the level the learner already earned, so
+// editing a definition can never roll a word back to level 1.
+func termLevel(status string, level *int, previous Term, exists bool) (int, error) {
 	if !termStatuses[status] {
 		return 0, failure(400, "词语状态无效")
 	}
 	if status == termStatusLearning {
-		if level == nil || *level == 0 {
-			return 1, nil
+		if level != nil && *level != 0 {
+			if *level < 1 || *level > termLevelMax {
+				return 0, failure(400, "学习等级须为 1～7")
+			}
+			return *level, nil
 		}
-		if *level < 1 || *level > termLevelMax {
-			return 0, failure(400, "学习等级须为 1～7")
+		if exists && previous.Status == termStatusLearning && previous.Level >= 1 {
+			return previous.Level, nil
 		}
-		return *level, nil
+		return 1, nil
 	}
 	if level != nil && *level != 0 {
 		return 0, failure(400, "只有学习中的词语可以设置等级")
@@ -193,6 +197,21 @@ func languageOf(tx *gorm.DB, owner int) (string, error) {
 	return space.Language, nil
 }
 
+// previousTerm reads the row this save is about to change under a lock, so the level and
+// the review schedule can be compared with what the learner already had.
+func previousTerm(tx *gorm.DB, owner int, language, term string) (Term, bool, error) {
+	var existing Term
+	err := tx.Clauses(clause.Locking{Strength: "UPDATE"}).
+		Where("owner_id = ? AND language = ? AND term = ?", owner, language, term).First(&existing).Error
+	if errors.Is(err, gorm.ErrRecordNotFound) {
+		return Term{}, false, nil
+	}
+	if err != nil {
+		return Term{}, false, err
+	}
+	return existing, true, nil
+}
+
 // saveTerm writes one identity with INSERT ... ON DUPLICATE KEY UPDATE: a repeated
 // save updates the same row instead of adding a second, conflicting record, and
 // two concurrent saves of the same word still leave exactly one.
@@ -219,9 +238,11 @@ func saveTerm(tx *gorm.DB, owner int, language, word string, fields TermFields,
 	if err := tx.Where("owner_id = ? AND language = ? AND term = ?", owner, language, row.Term).First(&stored).Error; err != nil {
 		return TermSave{}, err
 	}
-	// Saving a word also places it on the review schedule, so the loop from reading to
-	// reviewing needs no separate step.
-	if err := syncTermReview(tx, stored, now); err != nil {
+	// A saved word always owns a schedule row, but the date only moves for a new word or a
+	// real status/level change: editing a definition must not push a word out of today's
+	// queue (decision D2).
+	reschedule := !fields.Exists || fields.PreviousStatus != stored.Status || fields.PreviousLevel != stored.Level
+	if err := syncTermReview(tx, stored, reschedule, now); err != nil {
 		return TermSave{}, err
 	}
 	return TermSave{termView(stored), insert.RowsAffected == 1}, nil
@@ -268,12 +289,16 @@ func attachTerms(tx *gorm.DB, owner int, language string, tokens []TextToken) er
 	return nil
 }
 
-// TermFields carries the validated learner text and state into storage.
+// TermFields carries the validated learner text and state into storage, together with the
+// status and level the row had before this save.
 type TermFields struct {
-	Definition string
-	Examples   string
-	Status     string
-	Level      int
+	Definition     string
+	Examples       string
+	Status         string
+	Level          int
+	PreviousStatus string
+	PreviousLevel  int
+	Exists         bool
 }
 
 type TermInput struct {
@@ -306,7 +331,17 @@ func registerTermRoutes(v *gin.RouterGroup, protect func(bool, func(*gin.Context
 		if err != nil {
 			return nil, err
 		}
-		level, err := termLevel(input.Status, input.Level)
+		language, err := languageOf(tx, u.UserId)
+		if err != nil {
+			return nil, err
+		}
+		// The previous row decides whether a missing level keeps the earned one and whether
+		// the review date may move at all.
+		previous, exists, err := previousTerm(tx, u.UserId, language, normalizeWord(word))
+		if err != nil {
+			return nil, err
+		}
+		level, err := termLevel(input.Status, input.Level, previous, exists)
 		if err != nil {
 			return nil, err
 		}
@@ -314,11 +349,10 @@ func registerTermRoutes(v *gin.RouterGroup, protect func(bool, func(*gin.Context
 		if err != nil {
 			return nil, err
 		}
-		language, err := languageOf(tx, u.UserId)
-		if err != nil {
-			return nil, err
+		fields := TermFields{
+			Definition: definition, Examples: examples, Status: input.Status, Level: level,
+			PreviousStatus: previous.Status, PreviousLevel: previous.Level, Exists: exists,
 		}
-		fields := TermFields{Definition: definition, Examples: examples, Status: input.Status, Level: level}
 		return saveTerm(tx, u.UserId, language, word, fields, now())
 	}))
 	v.GET("/terms/:id", protect(false, func(c *gin.Context, tx *gorm.DB, u admin.SysUser) (any, error) {
diff --git a/server/app/lexgo/terms_test.go b/server/app/lexgo/terms_test.go
index 326048a..8bd9aeb 100644
--- a/server/app/lexgo/terms_test.go
+++ b/server/app/lexgo/terms_test.go
@@ -14,31 +14,44 @@ import (
 
 func TestTermLevelBoundary(t *testing.T) {
 	level := func(v int) *int { return &v }
+	// The stored state the save is about to change: a level-4 learning word.
+	stored := Term{Status: termStatusLearning, Level: 4}
 	cases := []struct {
-		status string
-		level  *int
-		want   int
-		ok     bool
+		status   string
+		level    *int
+		previous Term
+		exists   bool
+		want     int
+		ok       bool
 	}{
-		{termStatusNew, nil, 0, true},
-		{termStatusNew, level(0), 0, true},
-		{termStatusNew, level(1), 0, false},
-		{termStatusKnown, level(0), 0, true},
-		{termStatusKnown, level(3), 0, false},
-		{termStatusIgnored, level(0), 0, true},
-		{termStatusIgnored, level(-1), 0, false},
-		{termStatusLearning, nil, 1, true},
-		{termStatusLearning, level(0), 1, true},
-		{termStatusLearning, level(1), 1, true},
-		{termStatusLearning, level(7), 7, true},
-		{termStatusLearning, level(8), 0, false},
-		{termStatusLearning, level(-1), 0, false},
-		{"", nil, 0, false},
-		{"Learning", nil, 0, false},
-		{"deleted", nil, 0, false},
+		{termStatusNew, nil, Term{}, false, 0, true},
+		{termStatusNew, level(0), Term{}, false, 0, true},
+		{termStatusNew, level(1), Term{}, false, 0, false},
+		{termStatusKnown, level(0), stored, true, 0, true},
+		{termStatusKnown, level(3), stored, true, 0, false},
+		{termStatusIgnored, level(0), stored, true, 0, true},
+		{termStatusIgnored, level(-1), stored, true, 0, false},
+		{termStatusLearning, nil, Term{}, false, 1, true},
+		{termStatusLearning, level(0), Term{}, false, 1, true},
+		{termStatusLearning, level(1), stored, true, 1, true},
+		{termStatusLearning, level(7), stored, true, 7, true},
+		{termStatusLearning, level(8), stored, true, 0, false},
+		{termStatusLearning, level(-1), stored, true, 0, false},
+		{"", nil, stored, true, 0, false},
+		{"Learning", nil, stored, true, 0, false},
+		{"deleted", nil, stored, true, 0, false},
 	}
+	// A save that does not mention a level keeps the earned one instead of resetting it.
+	cases = append(cases, struct {
+		status   string
+		level    *int
+		previous Term
+		exists   bool
+		want     int
+		ok       bool
+	}{termStatusLearning, nil, stored, true, 4, true})
 	for _, tc := range cases {
-		got, err := termLevel(tc.status, tc.level)
+		got, err := termLevel(tc.status, tc.level, tc.previous, tc.exists)
 		if tc.ok && (err != nil || got != tc.want) {
 			t.Fatalf("%s/%v: got %d, %v; want %d", tc.status, tc.level, got, err, tc.want)
 		}