test(rectification): prove memoization in-process; tolerate golden float drift (BUG-733)

Replace whole-structure golden == with same-process cached vs None fallback equality, plus discrete-strict / measured-tolerance golden comparison. Do not rebuild the golden JSON.

BLOCKED.md

@@ -1,5 +1,10 @@
 # BLOCKED
 
+## 对话额度拆分（2026-09-16，分支 `codex/consultation-session-capacity-20260915`，BUG-732）
+
+- **无 Docker：** 本机 `docker` 不在 PATH。`npm run test:db` 与 `npm run db:migrate:check`（需要 `SCHEMA_DATABASE_URL` 连 Postgres）未跑。`frontend/tests/database-consultation-session-capacity.test.ts` 在无 Docker 时 skip，**不得写成通过**。替代证据是 SQL 静态合同：额度求和表达式不含 `thinkingText` / `thinkingSections`，物理上限为 `length(elem::text)`，算式写在迁移注释里。
+- **详情接口体积（5.3，已量，不改接口）：** 用三域本命夹具（每轮正文 4,000 汉字 + 思考 4,000 汉字 + 实测 `thinkingSections` + 典型三个 receipt）构造 `GET /api/sessions/[id]` 的 `{ session }` JSON。19 轮 UTF-8 **562,240 B（0.536 MiB）**；50 轮 **1,479,034 B（1.411 MiB）**，约为 2.63 倍。1.4 MiB 在 2 vCPU 上打开长会话会偏沉，分页 / 按需加载历史不在本单（任务书 §9）。PostgreSQL `length()` 按字节计，汉字正文 4,000 字 ≈ 12,000 字节，线上按字节撞 200,000 额度会早于「50 轮汉字」；50 轮数字是任务书字符口径对照。
+
 ## 只读页验收修复：镜像与浏览器（2026-09-16，分支 `codex/readonly-pages-fix-20260916`）
 
 - **无 Docker：** 无法验证 API 镜像里 `COPY vendor` 与容器内 `node --version`。不得写成通过。

CHANGELOG.md

@@ -1,5 +1,9 @@
 # 印度占星 Skill 更新日志
 
+## 2026-09-16 — 这段对话写满之前，思考过程不再占用额度
+
+普通咨询里，模型内部的思考过程不再算进「这段对话已写满」的额度。额度仍按你看得到的正文计算，能问的轮次大约从十几轮提到约五十轮。写满时还是请开一个新对话，提示不变。Skill 版本不变。
+
 ## 2026-09-16 — 对话写满后开新对话会静默带上之前的摘要
 
 普通咨询写满、点「开新对话」之后，服务端会把上一场的会话摘要拷进新对话，界面上看不出差别。进不了窗口的更早几轮，模型会看到省略说明，不再当成没发生过。Skill 版本不变。

docs/BUG_HISTORY.md

@@ -11380,3 +11380,35 @@
 - 相关记录：BUG-464、BUG-555
 - 复发自：无
 - 修复版本：待发布
+
+## BUG-732 | 对话额度把思考文本算进去，十几轮就「已写满」
+
+- 状态：resolved
+- 首次发现：2026-09-15
+- 最近更新：2026-09-16
+- 影响面：`append_consultation_question`、`POST /api/consult` 的 `session_full`、咨询会话详情 `GET /api/sessions/[id]`
+- 用户现象：普通咨询问大约十几轮就提示「这段对话已写满，开个新对话继续吧」。200 条消息那档永远碰不到。
+- 触发条件：继续往同一段咨询会话里发问；助手消息带有 `thinkingText` / `thinkingSections`。
+- 根因：BUG-464 立下的 200,000 字符上限身兼二职却两职都没做好。求和把用户读不到的 `thinkingText`、`thinkingSections` 算进去，却不算同样入库的 `techniqueTruth` / `workflowReceipt` / `agentExecutionReceipt`。这不是回归，是那条上限从一开始就混用了「对话有多长」和「这一行有多大」。
+- 修复：新迁移 `CREATE OR REPLACE` 该函数。对话额度仍是 200,000，只累加 `elem->>'text'`。另加物理上限 `sum(length(elem::text))`，算式 50 轮 ×（正文约 4,000 + 思考 4,000 + 分节 3,000 + 三个 receipt 约 3,000）≈ 700,000，取 1,000,000。两档都返回既有 `session_full`。签名、返回列、error_code、advisory lock、`request_id` 幂等、200 条上限、单条 16,000 字校验均未改。
+- 验证：`frontend/tests/consultation-session-capacity.test.ts` 锁定额度求和不含 `thinkingText` / `thinkingSections`、物理上限算式、签名与 `session_full`。`frontend/tests/database-consultation-session-capacity.test.ts` 用真实 Postgres 覆盖思考不占额度、短正文+大 receipt 撞物理上限、额度边界不先撞物理上限；本机无 Docker，该文件 skip，不得写成通过。既有幂等 / 满员用例未改。
+- 防复发：会话上限必须分成两条各司其职的口径：面向用户的对话额度只数用户读得到的正文；面向存储的物理上限必须把整条消息 JSON 算全。新增会存进 `messages` 的字段时，必须明确它进哪一条，不得默认落进对话额度。
+- 相关记录：BUG-464
+- 复发自：无
+- 修复版本：待发布
+
+## BUG-733 | 校正记忆化 golden 对全精度浮点做整体 `==`，跨机门禁靠运气绿
+
+- 状态：resolved
+- 首次发现：2026-09-15
+- 最近更新：2026-09-16
+- 影响面：`tests/test_rectification_engine_memoization.py`、`tests/golden/rectification_engine_memoization_v1.json`（文件未改）
+- 用户现象：用户看不见。门禁 glob `tests/test_rectification_*.py` 在 `CORE_PYTEST_TARGETS` 里，换一台机器或基础镜像会因 `score` / `margin_percent` 尾数差 1.0e-4～1.1e-3 而红，CI 绿只是碰巧和生成 golden 的那台舍入一致。
+- 触发条件：`test_score_candidates_matches_baseline_golden` 把 live payload 与 golden 做整体 `==`。
+- 根因：golden 存了全精度浮点并整体相等。这是 BUG-712 的同一形状。BUG-712 的防复发只落在 `ephemeris_events` 那一处（量化到 6 位再比），没有仓库级守卫，BUG-721 新写的 rectification golden 重蹈覆辙。
+- 修复：只改测试，不动 `scripts/`。主证据改成同进程差分：同一批 static contexts，A 带四层缓存、B 把 `ashtakavarga_result` / `shadbala_result` / `vimshottari_timeline` / `narayana_periods` 置 `None`，`compute_event_candidate_rows` 输出严格相等。golden 改成分档比较：离散字段严格相等；浮点用实测最大漂移 1.1e-3 推出的容差（绝对 2e-3 与相对 5e-4 取更宽者）。禁止调用 `write_golden()` 更新那份 JSON。
+- 验证：同进程 A/B 逐字相等，且 B 的 `calc_shadbala` 次数显著高于 A（本机 6 vs 0）。反向：把某一层缓存换成全零 shadbala 对象后不再相等。golden 分档比较通过；把 golden 浮点改 1e-2 必须红；把离散字段 `time` 改掉必须红。`pytest tests/test_rectification_engine_memoization.py` 与 `pytest tests/test_rectification_*.py` 0 failed。golden JSON `git diff` 无改动。
+- 防复发：任何 golden 比较都不得对浮点做整体 `==`；浮点必须量化或带容差，容差数值要有实测依据并写在注释里；能在同进程内做差分证明的命题，不得用跨机 golden 代替。
+- 相关记录：BUG-712、BUG-721
+- 复发自：BUG-712
+- 修复版本：待发布

docs/tasks/PROGRESS-consultation-session-capacity-20260915.md

@@ -0,0 +1,63 @@
+# PROGRESS · 对话上限一半被思考文本吃掉（2026-09-15）
+
+工作树：`.worktrees/consultation-session-capacity-20260915`  
+分支：`codex/consultation-session-capacity-20260915`  
+任务书基线：`6b3248bf`；开工时 `origin/staging` = **`11893c7f`**。  
+修复提交：`11829cf7`。  
+本机 Windows。无 Docker。
+
+未改 `frontend/src/app/api/consult/route.ts`、`page.tsx`、Skill。未 bump Skill。BUG 号核对：该 SHA 最大号仍是 **BUG-720**，本单用预占 **BUG-732**，无冲突。
+
+| 任务 | 状态 | 说明 |
+| --- | --- | --- |
+| 5.1 对话额度只数 `text` | 完成（运行时 DB 为环境缺口） | 新迁移 `20260916010000_consultation_session_capacity.sql` `CREATE OR REPLACE`；`v_chars` 只 `sum(length(elem->>'text'))` |
+| 5.2 物理上限 `length(elem::text)` | 完成（同上） | 算式见下；两档都返回 `session_full` |
+| 5.3 详情接口体积 | 完成（夹具量过） | 见下表；1.41 MiB 写入 `BLOCKED.md` 观察，本单不改接口 |
+| 5.4 BUG-732 | 完成 | 关联 BUG-464；不是回归 |
+
+## 实现要点
+
+- 只新增一份迁移。函数签名、返回列、`error_code` 取值、advisory lock、`request_id` 幂等、200 条上限、`char_length(v_text) > 16000` 均未改。
+- 物理上限算术（迁移注释原文）：50 rounds × (~4,000 body + ~4,000 thinkingText + ~3,000 thinkingSections + ~3,000 receipts) ≈ 700,000。Headroom → 1,000,000。
+- 实测 receipt（典型三域本命：12 行技法审计 + workflow + agentExecutionReceipt）JSON **1,738 字符 / 1,950 字节**，小于任务书 3,000，**不调高** 1,000,000。
+- 实测 `thinkingSections` JSON 字符：1 域 1,246、2 域 1,968、3 域 2,702（任务书 1,521 / 2,243 / 2,977，同量级，计划文案有漂移）。
+- 未删、未截断任何已存字段。
+
+## 详情 JSON 体积（5.3）
+
+夹具：每轮用户提问 100 汉字 + 助手正文 4,000 汉字 + 思考 4,000 汉字 + 三域 `natalConsultationThinkingPlan` + 上面那份典型 receipt。`GET /api/sessions/[id]` 的 `{ session }`（`sessionSelect` 含 `messages`）。
+
+| 口径 | 轮次 | UTF-8 字节 | MiB |
+| --- | ---: | ---: | ---: |
+| 改前（旧合并额度大约卡住的地方） | 19 | 562,240 | 0.536 |
+| 改后（任务书字符口径约 50 轮） | 50 | 1,479,034 | 1.411 |
+
+50 / 19 ≈ **2.63**。1.411 MiB 在 2 vCPU 上打开长会话会偏沉，观察项进 `BLOCKED.md`；分页不在本单。
+
+PostgreSQL `length()` 按字节，与 BUG-464 相同。4,000 汉字正文 ≈ 12,000 字节，按字节撞 200,000 会早于「50 轮汉字」。上表 50 轮是任务书字符口径对照，不是汉字会话在线上能存到的轮数。物理上限 1,000,000 字节在这组夹具上仍晚于对话额度触发。
+
+## 既有断言改动
+
+| 文件 | 原值 | 新值 | 原因 |
+| --- | --- | --- | --- |
+| `database-local-business.test.ts` 迁移已应用列表 | 最末 `20260915010000_rectification_touch_chat_session.sql` | 另加 `20260916010000_consultation_session_capacity.sql` 及「不得复制进 db/migrations」 | 新迁移必须被 runner 应用到 |
+
+未改幂等 / advisory lock / `session_missing` / 200 条满员三条既有用例。未改 `chat-session-authority.test.ts`（仍读 20260901010000 原文件，原文件未改）。
+
+## 测试
+
+| 命令 | 结果 |
+| --- | --- |
+| `./node_modules/.bin/tsc --noEmit` | 0 错 |
+| `npx tsx --test tests/consultation-session-capacity.test.ts` | 6 pass / 0 fail |
+| `npx tsx --test tests/consultation-*.test.ts tests/chat-session-*.test.ts` | 220 tests / 208 pass / **12 fail**：全部是 Windows `SKILL.md` symlink `EPERM`（`consultation-agentic-runtime`、`consultation-birth-time-mode`、`consultation-workflow-request` 整文件，以及 methodology 读 live skill 的 9 条）。与本单无关，既有环境缺口。本单 6 条与 `chat-session-*` 全绿 |
+| `npx eslint`（本单三个测试文件 `--max-warnings 0`） | 0 error / 0 warning |
+| `npm run lint` | **0 error** / 119 warning（既有，未动） |
+| `npm run db:migrate:check` | **blocked**：`SCHEMA_DATABASE_URL is required` |
+| `npm run test:db` | **blocked**：39 tests / 7 pass / 27 fail / 5 skip。失败均为 `spawnSync docker ENOENT` 或 Windows symlink EPERM。本单 `database-consultation-session-capacity` **skip**（`docker unavailable on this host`），不得写成通过 |
+| 全量 `npx tsx --test tests/*.test.ts` | 3089 tests / 3001 pass / 73 fail / 15 skip。失败为无 Docker `ENOENT`、Windows `SKILL.md` symlink `EPERM`、以及 `G:\\G:\\` 路径加倍，与既有环境缺口同类。本单 6 条静态合同在该套件里通过；DB 用例 skip |
+
+## 收尾限制
+
+- 运行时 Postgres 合同本机未跑。有 Docker 的验收机必须跑 `npm run test:db`。
+- 详情接口未分页。50 轮夹具 1.41 MiB，见 `BLOCKED.md`。

docs/tasks/PROGRESS-rectification-engine-memoization-fix-20260915.md

@@ -0,0 +1,64 @@
+# PROGRESS · 等价 golden 跨机不稳（2026-09-15 / 执行 2026-09-16）
+
+- 执行分支：`codex/rectification-engine-memoization-fix-20260915`
+- 工作树：`.worktrees/rectification-engine-memoization-fix-20260915`
+- 任务书：`docs/tasks/TASK-rectification-engine-memoization-fix-20260915.md`
+- 开工 `HEAD` = `origin/staging` = `51a65d92`（用户核对时是 `7227b1ed`，其后合入了 BUG-732）
+- BUG：`BUG-733`（复发自 BUG-712；落库时最大号 732）
+
+## 开工前置
+
+| 项 | 结果 |
+| --- | --- |
+| `git status -sb` 第一行 | `## codex/rectification-engine-memoization-fix-20260915...origin/staging` |
+| `HEAD` | `51a65d92` |
+| BUG_HISTORY | 最大号 **BUG-732**。本单预占 **BUG-733**。读完 BUG-712 / BUG-721 |
+| 本机 Python | 主仓 `.venv` 是 25 字节指针文件。本 worktree 用 Anaconda 3.11.7 `--system-site-packages` 建真实 venv，复用 `swisseph 20230604`，未升级依赖 |
+| 先复现 golden | `test_score_candidates_matches_baseline_golden` **本机绿**。live 与 golden 逐字相同：`score` 8.6274 / 8.6227 / 8.1701，`margin_percent` 5.3006。验收机上的 1.1e-3 漂移在这台 Windows / Anaconda 3.11.7 上复现不出来——golden 就是同类机器生成的。仍按任务书改比较方式，不调用 `write_golden()` |
+
+未改 `docs/tasks/README.md` 状态列。未推 staging。未改任何 `scripts/` 实现。未改 `tests/golden/rectification_engine_memoization_v1.json`。
+
+## 做了什么
+
+| 条 | 做法 |
+| --- | --- |
+| 5.1 | 新增 `test_cached_static_context_matches_uncached_fallback`：同一请求、同一进程，A=带四层缓存的 static contexts，B=四键置 `None`。`compute_event_candidate_rows` 输出 `==`。调用计数：A 的 `calc_shadbala` = 0，B = 6（3 候选 × 2 事件）。反向：`test_poisoned_static_cache_diverges_from_live_rows` 把第一候选的 `shadbala_result` 换成全零对象，分数从 12.5125 变成 12.6125，断言不再相等 |
+| 5.2 | `test_score_candidates_matches_baseline_golden` 改成分档比较。整数 / 布尔 / 字符串 / 列表身份严格相等；浮点容差 `max(2e-3, 5e-4 * \|expected\|)`，注释写明来自实测最大漂移 1.1e-3。反向：`score += 1e-2` 红；`time = "99:99"` 红 |
+| 5.3 | 未调用 `write_golden()`。`git diff` 里该 JSON 无改动 |
+| 5.4 | 六份检查结论见下表。同类问题只记录不修 |
+| 5.5 | `docs/BUG_HISTORY.md` 新增 BUG-733，复发自 BUG-712。证据是 5.1 同进程差分 + 两次反向验证，不是「现在绿了」 |
+
+## 5.4 `tests/golden/` 检查
+
+| 文件 | 全精度浮点整体 `==` | 在哪个测试里比 | 怎么比 |
+| --- | --- | --- | --- |
+| `consultation_contract_keypaths_v1.json` | 无 | `tests/test_consultation_contract_golden.py` | 只存 keypath 与 JSON 类型（`kind=keypath_types_only`）。值不入库。live 是 keypath 超集 + 类型相同 |
+| `ephemeris_events_raman_20260915_90d.json` | 无（BUG-712 已改） | `tests/test_ephemeris_events.py` | `longitude` / `speed_longitude` 量化到 6 位再 `==`；`kind` / `date` / `body` / 星座字段严格相等。文件里仍存全精度，但测试不再整体 `==` |
+| `golden_cases.json` | 无 | `tests/test_cli_smoke.py::test_full_reading_golden_cases_cover_user_ready_output` → `tests/run_golden_cases.py` | 路径非空、模块计数、整数不变量 `ashtakavarga_sav_total == 337`、日期覆盖。不是把整份引擎浮点输出整体 `==` |
+| `qizheng_stem_branch_19900409.json` | 无（比较已量化） | `tests/test_qizheng_chart_engine.py` | 离散字段（宫、宿、ketuMode）严格相等；`siderealLon` / `mansionDegree` 用 `round(..., 2)`。**文件本身仍存全精度浮点**（如 `177.94225865978694`）。若有人新写 `assert result == golden` 会变成 BUG-712 同类。本单只记录不修 |
+| `rectification_engine_memoization_v1.json` | **本单之前有** | `tests/test_rectification_engine_memoization.py::test_score_candidates_matches_baseline_golden` | 改前：`candidate_scores` 与 `decision_receipt` 整体 `==`。改后：离散严格、浮点容差。文件未重建 |
+| `upstream_sync2/` | 无整体浮点 `==` | 见分项 | `relationship_einstein.json`：`test_upstream_sync2_goldens.py` 只比键与类型。`consultation_marriage_evidence_snapshot.json`：shape。`consultation_marriage_1990_fictional.json`：`test_relationship_event_class_evidence.py` 比离散 Dasha 字符串。`pl9_time_system_table.baseline.md`：`test_full_reading_conditional_dashas.py` 比对 markdown 表行文本（前 5 行 `==`），不是 JSON 浮点。`pl9_time_system_table.md` 与 `pl9_export_timing.json`（含 `seconds: 8.647`）**没有任何测试读取** |
+
+同类观察（不修）：qizheng golden 文件仍存全精度；ephemeris golden 文件仍存全精度（测试已量化）；`upstream_sync2/pl9_export_timing.json` 是无引用的浮点秒数。
+
+## 测试
+
+| 套件 | 结果 |
+| --- | --- |
+| `tests/test_rectification_engine_memoization.py` | **14 passed**（原 10 + 新增 4） |
+| `tests/test_rectification_*.py` | **186 passed, 0 failed**（14.42s） |
+
+命令使用 `.\.venv\Scripts\python.exe`（任务书写的 `.venv/bin/python` 在 Windows 不存在）。
+
+未改任何既有断言的期望值。新增 4 条：同进程差分、缓存毒化反向、浮点 1e-2 反向、离散字段反向。原 golden 断言从整体 `==` 改成分档，不是放宽离散字段。未跑 `run_quality_gate.py --profile quick`：该 profile 在 pytest 之后会跑 `npm test`，本机 Windows 缺 `tsx` PATH（与 BUG-721 进度记录同一缺口），本单未改前端。
+
+## CHANGELOG
+
+未改。纯测试比较方式，用户可见打分不变。
+
+## 偏离
+
+- 任务书写 `.venv/bin/python`，本机换成 `.\.venv\Scripts\python.exe`。
+- 本机未能复现验收机上的 golden 红灯（舍入与生成机一致）。仍改比较方式，并保留验收机测到的 1.1e-3 作为容差依据。
+- 未 `git push origin HEAD:staging`。
+- 5.4 按让步只检查并记录，未加仓库级自动扫描守卫。

docs/tasks/README.md

@@ -240,9 +240,10 @@
 | `TASK-rectification-request-dossier-cache-20260915.md` | — | **低风险单，串行在 failure-attribution 之后（同改 `route.ts`）**：一轮 Agent 对话实测取 3.44 次整份 Case 档案（点选题 2.07 次），全仓约 40 个调用点、请求内零缓存；档案是「最近 50 轮 turns + 全部 evidence + 合成收据」的大 jsonb。做法是包装 `accounting` 客户端做**写即失效**的请求作用域缓存（两个只读投影命中缓存，其余任何 RPC 先清空再转发），**零调用点改动**。不得做成「请求内只读一次」——档案在请求内会变。BUG 段 726 | 待领取 | — |
 | `TASK-consultation-external-evidence-cache-20260915.md` | `PROGRESS-consultation-external-evidence-cache-20260915.md` | **普通聊天性能单（Python；2026-09-15 产品拍板改为排在 api-server-decomposition 之前）**：每轮每域同步等外网，cProfile 前三名全是 `api.vedastro.org` 的 HTTPS 往返（0.801 + 0.786 + 0.206 s），本地 swisseph 只有 0.022 s。三个护栏数字凑不齐：前台等 1.5 s、后台跑 8 s、线程池只有 2 个 worker，且超时**不 cancel** → 每 4 秒一轮就长期饱和，之后每轮白等再拿 `official_blocked`（BUG-727）。另 `western_evidence_packet` 122 KB 前端零读取点（BUG-728）。**产品定案**：按「出生数据+岁差+交点+UTC 日期」缓存（与引擎 `_official_snapshot_reference_date` 同键，否决自定 TTL），同日 0 等待 / 跨日先用旧的（≤7 天）后台刷新 / `daily_starlanguage` 要求当天 / 冷启动才走 1.5 s。**不许「干脆不调」——那会重开 BUG-301。** 另含 staging 单域耗时实测单（代码注释里的 21 s 与本机 0.5 s 差 40 倍，三域上限就是从它推的）。BUG 段 727–728 | 待验收 | `codex/consultation-external-evidence-cache-20260915` |
 | `TASK-consultation-context-memory-20260915.md` | `PROGRESS-consultation-context-memory-20260915.md` | **记忆三缺口（TS，可并行）**：历史超预算时从最老整轮丢弃，`droppedCount` 算了却**全仓零读取点**，模型不知道少看了几轮——单条截断有「省略 N 字」标记，整轮丢弃没有（BUG-729，BUG-555 防复发只写了「头部截断」所以漏网）；写摘要阈值写死 16,000，历史预算却是 `clamp((窗口−60k)×1.5, 4k, 40k)`，窗口 < **70,667** 时预算低于阈值 → 每轮静默丢（BUG-730，后台上架中等窗口模型即触发）；写满时服务端存着摘要，`continueInNewChat` 只带问题不带摘要，而 `context_summary` 根本不在任何会话接口的列里（BUG-731）。**产品定案：静默继承**，且摘要文本永远不许由客户端提供（`chatSessionCreateSchema` 只收来源会话 uuid）。BUG 段 729–731 | 待验收 | `codex/consultation-context-memory-20260915` |
-| `TASK-consultation-session-capacity-20260915.md` | — | **对话上限单（一份迁移，可并行；不碰 route.ts）**：`append_consultation_question` 的 200,000 字符额度里，`thinkingText`(≤4,000) + `thinkingSections`(实测 1,521/2,243/2,977) 占一半以上，而 `techniqueTruth`/`workflowReceipt`/`agentExecutionReceipt` 照样入库却不计入——同一条上限身兼二职且两职都没做好，约 **19 轮** 就「已写满」（200 条那档永远碰不到）。**产品定案：思考文本不计入**，额度只数用户读得到的正文（约 19 → 约 50 轮），另设一条按 `length(elem::text)` 把全部字段算全的物理上限（算式取 1,000,000，写进迁移注释）护住数据库行；两档都返回同一个 `session_full`。保留 advisory lock / 幂等 / 满员拒绝（BUG-464 防复发）。BUG 段 732 | 待领取 | — |
+| `TASK-consultation-session-capacity-20260915.md` | `PROGRESS-consultation-session-capacity-20260915.md` | **对话上限单（一份迁移，可并行；不碰 route.ts）**：`append_consultation_question` 的 200,000 字符额度里，`thinkingText`(≤4,000) + `thinkingSections`(实测 1,521/2,243/2,977) 占一半以上，而 `techniqueTruth`/`workflowReceipt`/`agentExecutionReceipt` 照样入库却不计入——同一条上限身兼二职且两职都没做好，约 **19 轮** 就「已写满」（200 条那档永远碰不到）。**产品定案：思考文本不计入**，额度只数用户读得到的正文（约 19 → 约 50 轮），另设一条按 `length(elem::text)` 把全部字段算全的物理上限（算式取 1,000,000，写进迁移注释）护住数据库行；两档都返回同一个 `session_full`。保留 advisory lock / 幂等 / 满员拒绝（BUG-464 防复发）。BUG 段 732 | 待验收 | `codex/consultation-session-capacity-20260915`（BUG-732）；`test:db` 环境缺口 |
 | `TASK-freeze-metric-change-20260915.md` | — | **规则单（后面两单的前置，无 BUG 号）**：两条增长冻结余量都用完（`page.tsx` 1,951/1,951 余 **0**；`jyotish_api_server.py` 11,334/11,363 余 **29**），冻结从「逼新代码往外走」退化成「拦路」。实证：`page.tsx` 行数砍 59% 但 `Home()` 的 `useState` 从 56 涨到 **66**（拆的是代码不是状态）；api server **225 个类方法只有 12 处真碰 HTTP 上下文**，4 处 `__new__` 伪造空壳就是这么来的。**产品拍板换口径**：主门改成「`Home()` 的 useState/useRef 不得增长」与「类方法数 + `__new__` 计数不得增长」，行数降级为粗护栏；**同时推翻 §6「参数式 hook 内部保持 0 个 React hook」**（那正是状态搬不走的原因）。改 `AGENTS.md` §6 + 两个合同测试，不碰业务代码 | 待领取 | — |
 | `TASK-home-state-lowering-20260915.md` | — | **page.tsx 状态下沉第一簇（串行在 freeze-metric-change + C2 + R3 之后）**：66 个 state 里 `rectification*` 占 **15** 个，而它们服务的 `<ConversationalBirthTimeRectification>` 本来就是 `dynamic()` 懒加载子树、挂着 24 个 props；`useRectificationSurface` 要解构约 56 个参数。把这簇搬进子树，`Home()` 的 useState 从 66 降到 ≤ 53。**零行为变化**；第一步必须先把 15 个逐个分类（只服务子树 / 外壳也要读）。产品否决了 Context Provider 与外部 store 两条路。不占 BUG 号 | 待领取 | — |
+| `TASK-rectification-engine-memoization-fix-20260915.md` | — | **验收修复单（只改测试，一行实现不许动）**：BUG-721 的实现**等价性成立**（我在改前 `6b3248bf` / 改后 `e4788dfc` 同机跑同一 payload，`candidate_scores` 逐字相同），9 条计数断言全过；但等价 golden 在本机复现不出来——4 处浮点尾数差（score 1.0e-4 ×2、`margin_percent` 1.1e-3 ×2）。**复发自 BUG-712**（「不得对全精度浮点做整体 `==`」，那一单只落在 ephemeris 一处）。而 `tests/test_rectification_*.py` 在 `CORE_PYTEST_TARGETS` 里，**staging 门禁靠机器舍入碰巧一致才是绿的**。修法：主证据换成**同进程差分**（把 static context 的四个缓存键置 `None` 即可回退旧路径，A/B 严格相等），golden 降为离散字段严格相等 + 浮点带容差（容差按实测 1.1e-3 推）；**禁止重建 golden 来「修」**。另含六份 golden 的仓库级排查。BUG-733 | 待领取 | — |
 
 ## 命名与归档
 

docs/tasks/TASK-rectification-engine-memoization-fix-20260915.md

@@ -0,0 +1,162 @@
+# TASK · 验收修复单：等价 golden 跨机不稳，门禁靠运气绿
+
+- 日期：2026-09-15
+- 基线 commit：`origin/staging` @ `e4788dfc`（实现落在 `53a37ce9`）
+- 执行分支：`codex/rectification-engine-memoization-fix-20260915`
+- 来源：Claude 对 `TASK-rectification-engine-memoization-20260915`（BUG-721）的验收
+- 落点：`tests/test_rectification_engine_memoization.py`、必要时 `tests/golden/`、`docs/BUG_HISTORY.md`
+- 规模：只改测试。**一行实现代码都不许动。**
+
+---
+
+## 1. 验收结论摘要
+
+BUG-721 的**实现是对的，等价性成立**，我独立复核过。门禁实测（基线 `11893c7f` → 头 `e4788dfc`，含 R1/R2/R3 三个提交）：
+
+| 项 | 结果 |
+| --- | --- |
+| `tsc --noEmit` | 0 错 |
+| `npm run lint` | **0 error** / 119 warning（全部既有） |
+| 全量前端套件 | 基线 3291（fail 31）→ 头 3304（fail 31），**失败清单逐条一致** |
+| `next build` | exit 0；`/` 仍 `○ Static` |
+| 首屏 JS gzip-9 | 基线 130,872 B → 头 130,872 B，**0.00%**（两侧同一种量法） |
+| `tests/test_rectification_*.py` | **1 failed, 181 passed** |
+
+唯一那条失败就是本单要修的：`test_score_candidates_matches_baseline_golden`。
+
+**它不是实现改坏了。** 我在改前（`6b3248bf`，与 golden 自称的 `a8d29d1b` 代码相同）和改后（`e4788dfc`）两个工作树上跑同一份 payload：
+
+| | 12:00 | 12:01 | 12:02 |
+| --- | ---: | ---: | ---: |
+| 本机 **改前** | 8.6273 | 8.6226 | 8.1701 |
+| 本机 **改后** | 8.6273 | 8.6226 | 8.1701 |
+
+**改前改后逐字相同。** 本单的其余 9 条断言（shadbala / ashtakavarga / vimshottari / narayana 各等于候选分钟数；过境盘等于去重后事件日期数；探针默认 1 次、refresh 时 2 次）全部通过。
+
+## 2. 事故实证
+
+失败来自 golden 文件与**任何一台机器**的实际输出对不上：
+
+| 字段 | 本机（改前＝改后） | `tests/golden/rectification_engine_memoization_v1.json` | 差 |
+| --- | ---: | ---: | ---: |
+| `candidate_scores[0].score` | 8.6273 | 8.6274 | 1.0e-4 |
+| `candidate_scores[1].score` | 8.6226 | 8.6227 | 1.0e-4 |
+| `decision_receipt.margin_percent` | 5.2995 | 5.3006 | 1.1e-3 |
+| `decision_receipt.gates.diagnostic_quality.margin_percent` | 5.2995 | 5.3006 | 1.1e-3 |
+
+一共 4 处，全是浮点尾数。`candidate_scores[*].score` 本身已经被引擎 `round(..., 4)` 过，差值正好是那一位的 1 ulp——也就是求和顺序/libm 舍入的跨机差异，不是算法差异。
+
+**后果是实打实的**：`tests/test_rectification_*.py` 这个 glob 在 `scripts/run_quality_gate.py` 的 `CORE_PYTEST_TARGETS` 里（注释原文：「Auto staging gate is `--profile quick`…so a stale window_scan assertion in this glob stayed red on origin/staging until listed here」）。也就是说 **staging 自动门禁会跑这条断言**。今天它在 CI 上绿，只是因为那台机器的舍入和执行方的机器一致；换一台机器、换一个基础镜像、换一次 libm 版本就会红。**这条门禁现在是靠运气绿的。**
+
+## 3. 根因
+
+golden 存了全精度浮点并做整体 `==`。这正是 **BUG-712** 的形状，那条记录的防复发一字不差地写着：
+
+> 不得对全精度浮点做整体 `==`。测试不得写 golden。
+
+BUG-712 修的是 `ephemeris_events` 的 golden（`longitude` / `speed_longitude` 量化到 6 位再比，非浮点字段保持严格相等）。那一单的范围写死在星历端点上，新写的这份 rectification golden 是同样的写法，没被扫到。
+
+更深一层：**golden 本来就不是证明「记忆化没改结果」的合适工具。** golden 证明的是「今天的输出等于某台机器某一天的输出」，中间夹了一个与被测命题无关的变量（机器）。被测命题其实可以在**同一个进程里**证明——见 §5.1。
+
+## 4. 决策记录
+
+产品 2026-09-15 授权本单：
+
+1. **不回滚 BUG-721 的实现。** 等价性已由改前/改后同机对比独立证实，实现留在 staging 上。
+2. **本单只改测试。** 一行 `scripts/` 下的实现代码都不许动——不得为了让断言通过去改业务代码（这是产品的既定偏好）。
+3. **不得用「重新生成 golden」来修。** 在本机重跑一次 `write_golden()` 能让测试变绿，但那只是把不稳定性换个方向藏起来，下一台机器照样红。**这条是硬红线，见 §5.3。**
+4. **防复发升级为仓库级。** BUG-712 的那句话从单点措施提升成对所有 golden 生效的规则，并配一条能自动发现违例的守卫。
+
+## 5. 任务分解
+
+### 5.1 把主证据换成同进程差分（首要，也是本单真正的价值）
+
+实现已经把四层不变量放进 static context 的具名键（`ashtakavarga_result` / `shadbala_result` / `vimshottari_timeline` / `narayana_periods`），而每个消费点都有 `if result is None:` 的回退分支。**这意味着记忆化可以在同一个进程里关掉**：把 context 里那四个键置 `None`，消费点就会退回逐次计算的老路径。
+
+新增一条断言：同一份请求、同一个进程，
+- A：正常的 static contexts（带四层缓存）
+- B：把那四个键全部置 `None` 的同一批 contexts（回退到逐次计算）
+
+断言 `compute_event_candidate_rows` 在 A / B 下的输出**严格逐字相等**。
+
+这条证明不经过任何跨机变量，是比 golden 强得多的等价证据。
+
+- 验收：新断言通过；人为把某一层的缓存值替换成错误对象后必须红（贴反向验证，证明这条断言不是恒为真）。
+- 验收：A / B 两条路径确实走了不同分支（用调用计数确认 B 的 `calc_shadbala` 次数显著高于 A），否则等于什么都没测。
+
+### 5.2 golden 断言改成分档比较
+
+保留 golden（它仍有价值：能发现跨版本的大幅漂移），但改比较方式：
+
+| 字段类别 | 比较方式 |
+| --- | --- |
+| 离散字段（`time`、`supporting_event_ids`、`conflicting_event_ids`、各 gate 布尔、`overall_confidence`、`representative_time`、排序） | **严格相等** |
+| 浮点字段 | 容差比较 |
+
+容差必须由实测漂移推出来，不得拍脑袋：本次实测最大漂移 **1.1e-3**（`margin_percent`）。取一个有余量但仍能发现真回归的值（建议绝对 2e-3 与相对 5e-4 取更宽者），并把「这个数怎么来的」写进测试注释。
+
+- 验收：把 golden 里任意一个浮点改动 1e-2，断言必须红（贴反向验证）。
+- 验收：把 golden 里任意一个离散字段改掉，断言必须红。
+- 验收：`.venv/bin/python -m pytest tests/test_rectification_*.py` 在本机 **0 failed**。
+
+### 5.3 不得重建 golden
+
+`write_golden()` 这个辅助函数留着没问题，但**本单不得调用它更新那份 golden 文件**。若执行方认为必须重建，要在进度记录里写明理由并说明为什么不是在掩盖跨机不稳——默认答案是「不重建」。
+
+- 验收：`git diff` 里 `tests/golden/rectification_engine_memoization_v1.json` **无改动**（若有改动，按上一条给出理由）。
+
+### 5.4 仓库级防复发守卫
+
+`tests/golden/` 下现在共 6 份：`consultation_contract_keypaths_v1.json`、`ephemeris_events_raman_20260915_90d.json`、`golden_cases.json`、`qizheng_stem_branch_19900409.json`、`rectification_engine_memoization_v1.json`、`upstream_sync2/`。
+
+- 逐份检查有没有「全精度浮点 + 整体 `==`」的比较方式，结果列进进度记录（每份写明：有/无、在哪个测试里比的、怎么比的）。
+- 发现同类问题的，本单**只记录不修**（各自另开单），除非改动小到一眼可见。
+
+- 验收：六份的检查结论在进度记录里，一份不漏。
+
+### 5.5 Bug 历史
+
+同一变更内写进 `docs/BUG_HISTORY.md`，预占 **BUG-733**（721–732 已被两轮审计的七单占用，避让）。必须写明：
+
+- **复发自 BUG-712**，并说明为什么没拦住：BUG-712 的防复发只落在 `ephemeris_events` 那一处，没有仓库级守卫，新写的 golden 重蹈覆辙。
+- 状态可写 `resolved`，但**证据必须是 §5.1 的同进程差分 + §5.2 的两次反向验证**，不能只写「现在绿了」。
+- 防复发升级为：**任何 golden 比较都不得对浮点做整体 `==`；浮点必须量化或带容差，容差数值要有实测依据并写在注释里；能在同进程内做差分证明的命题，不得用跨机 golden 代替。**
+
+## 6. 让步顺序
+
+1. 5.1 **不得砍**——它才是这一单的意义；没有它，5.2 只是把红灯调成绿灯。
+2. 5.2 必须做。
+3. 5.4 可以只做「检查并记录」，修留到后续单。
+4. 5.3、5.5 不得砍。
+
+## 7. 开工前置命令
+
+```bash
+git fetch origin --prune
+git worktree add -b codex/rectification-engine-memoization-fix-20260915 \
+  .worktrees/rectification-engine-memoization-fix-20260915 origin/staging
+cd .worktrees/rectification-engine-memoization-fix-20260915
+git status -sb | head -1
+# 先复现：这一条应当是红的
+.venv/bin/python -m pytest tests/test_rectification_engine_memoization.py -q
+```
+
+开工前必读：`docs/BUG_HISTORY.md` 的 **BUG-712**（同形态前例与它的修法）与 **BUG-721**（本单要保护的那次改动）。
+
+验收命令：
+
+```bash
+.venv/bin/python -m pytest tests/test_rectification_engine_memoization.py
+.venv/bin/python -m pytest tests/test_rectification_*.py          # 门禁 glob，须 0 failed
+.venv/bin/python scripts/run_quality_gate.py --profile quick
+```
+
+## 8. BUG 编号起点
+
+基线 `e4788dfc` 上 `docs/BUG_HISTORY.md` 最大号为 **BUG-725**。721–732 已被两轮审计七单预占（726–732 尚未落库），本单避让，预占 **BUG-733**。开工时核对实际最大号。
+
+## 9. 不在本单范围
+
+- BUG-721 的实现（已验收等价，不动）
+- 其余五份 golden 的修复（5.4 只检查并记录）
+- `CORE_PYTEST_TARGETS` 的构成（`tests/test_rectification_*.py` 这条 glob 保留，本单是让它变得可靠，不是把它摘掉）

frontend/supabase/migrations/20260916010000_consultation_session_capacity.sql

@@ -0,0 +1,135 @@
+-- BUG-732: split append_consultation_question into two caps.
+-- Conversation quota (200,000) counts only user-visible `text`.
+-- thinkingText / thinkingSections stay stored but must not enter that sum.
+-- Physical cap counts the whole message JSON, including receipts.
+-- Arithmetic: 50 rounds × (~4,000 body + ~4,000 thinkingText + ~3,000
+-- thinkingSections + ~3,000 receipts) ≈ 700,000. Headroom → 1,000,000.
+-- Signature, return columns, error_code values, advisory lock, request_id
+-- idempotency, 200-message cap, and 16,000-char question check are unchanged
+-- (BUG-464). CREATE OR REPLACE is backward compatible with deployed callers.
+
+begin;
+
+do $migration$
+begin
+  if current_user <> 'schema_owner' then
+    raise exception 'consultation_session_capacity_requires_schema_owner'
+      using errcode = '42501';
+  end if;
+end
+$migration$;
+
+create or replace function public.append_consultation_question(
+  p_user_id uuid,
+  p_request_id text,
+  p_session_id uuid,
+  p_question_message jsonb
+)
+returns table(success boolean, error_code text)
+language plpgsql
+security definer
+set search_path = ''
+as $$
+declare
+  v_session public.chat_sessions%rowtype;
+  v_message jsonb;
+  v_text text;
+  v_request_id text;
+  v_count integer;
+  v_chars integer;
+  v_new_chars integer;
+  v_physical_chars bigint;
+  v_new_physical bigint;
+begin
+  v_request_id := btrim(coalesce(p_request_id, ''));
+  if p_user_id is null or p_session_id is null or v_request_id = '' then
+    return query select false, 'invalid_request'::text;
+    return;
+  end if;
+  if jsonb_typeof(p_question_message) <> 'object'
+    or p_question_message->>'role' <> 'user' then
+    return query select false, 'invalid_question_message'::text;
+    return;
+  end if;
+  v_text := btrim(coalesce(p_question_message->>'text', ''));
+  if v_text = '' or char_length(v_text) > 16000 then
+    return query select false, 'invalid_question_message'::text;
+    return;
+  end if;
+
+  perform pg_advisory_xact_lock(hashtextextended(p_user_id::text || ':' || v_request_id, 0));
+
+  select session.* into v_session
+  from public.chat_sessions as session
+  where session.id = p_session_id
+    and session.user_id = p_user_id
+    and session.session_type = 'consultation'
+  for update;
+
+  if not found then
+    return query select false, 'session_missing'::text;
+    return;
+  end if;
+
+  if exists (
+    select 1
+    from jsonb_array_elements(coalesce(v_session.messages, '[]'::jsonb)) as elem
+    where elem->>'requestId' = v_request_id
+  ) then
+    return query select true, null::text;
+    return;
+  end if;
+
+  v_message := p_question_message || jsonb_build_object('requestId', v_request_id);
+
+  v_count := jsonb_array_length(coalesce(v_session.messages, '[]'::jsonb));
+  select
+    coalesce(sum(length(coalesce(elem->>'text', ''))), 0),
+    coalesce(sum(length(elem::text)), 0)
+  into v_chars, v_physical_chars
+  from jsonb_array_elements(coalesce(v_session.messages, '[]'::jsonb)) as elem;
+
+  v_new_chars := length(v_text);
+  v_new_physical := length(v_message::text);
+  if v_count >= 200
+    or (v_chars + v_new_chars) > 200000
+    or (v_physical_chars + v_new_physical) > 1000000 then
+    return query select false, 'session_full'::text;
+    return;
+  end if;
+
+  update public.chat_sessions as session
+  set messages = coalesce(session.messages, '[]'::jsonb) || jsonb_build_array(v_message),
+      title = case
+        when coalesce(btrim(session.title), '') in ('', '新对话') then
+          case
+            when char_length(v_text) > 14 then left(v_text, 14) || '…'
+            else v_text
+          end
+        else session.title
+      end,
+      updated_at = clock_timestamp()
+  where session.id = p_session_id
+    and session.user_id = p_user_id
+    and session.session_type = 'consultation';
+  if not found then
+    return query select false, 'session_missing'::text;
+    return;
+  end if;
+
+  return query select true, null::text;
+end;
+$$;
+
+revoke all on function public.append_consultation_question(uuid, text, uuid, jsonb)
+  from public, anon, authenticated;
+grant execute on function public.append_consultation_question(uuid, text, uuid, jsonb)
+  to service_role;
+do $$ begin
+  if exists(select 1 from pg_roles where rolname = 'admin_runtime') then
+    grant execute on function public.append_consultation_question(uuid, text, uuid, jsonb)
+      to admin_runtime;
+  end if;
+end $$;
+
+commit;

frontend/tests/consultation-session-capacity.test.ts

@@ -0,0 +1,288 @@
+import assert from "node:assert/strict";
+import { existsSync, readdirSync, readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import test from "node:test";
+
+import { natalConsultationThinkingPlan } from "../src/lib/consultation-thinking-plan.ts";
+
+const migrationsDir = new URL("../supabase/migrations/", import.meta.url);
+const originalSql = readFileSync(
+  new URL("../supabase/migrations/20260901010000_append_consultation_question.sql", import.meta.url),
+  "utf8",
+);
+const sql = readFileSync(
+  new URL("../supabase/migrations/20260916010000_consultation_session_capacity.sql", import.meta.url),
+  "utf8",
+);
+const itemRoute = readFileSync(
+  new URL("../src/app/api/sessions/[id]/route.ts", import.meta.url),
+  "utf8",
+);
+const consultRoute = readFileSync(
+  new URL("../src/app/api/consult/route.ts", import.meta.url),
+  "utf8",
+);
+
+const SESSION_SELECT =
+  "id,title,theme,model_id,messages,session_type,rectification_case_id,chart_profile_id,chart_profile_name,chart_profile_role,updated_at,pinned,archived_at";
+
+function functionBody(source: string): string {
+  const match = source.match(/as \$\$\r?\n([\s\S]*?)\r?\n\$\$;/);
+  assert.ok(match?.[1], "function body must be present");
+  return match[1];
+}
+
+function withoutLineComments(source: string): string {
+  return source.replace(/--[^\n]*/g, "");
+}
+
+function utf8Bytes(value: string): number {
+  return Buffer.byteLength(value, "utf8");
+}
+
+function typicalReceipts(domains: readonly string[]) {
+  const workflowReceipt = {
+    route: "multi-domain",
+    status: "ready",
+    preciseTiming: "allowed",
+    missingLayers: [] as string[],
+    domains,
+  };
+  const agentExecutionReceipt = {
+    runId: "run-typical",
+    runtime: "mastra-agentic" as const,
+    skill: {
+      name: "jyotish-vedic-astrology" as const,
+      loaded: true,
+      version: "1.0.0",
+      referenceReads: 2,
+      methodologySections: 4,
+    },
+    steps: [
+      { sequence: 1, kind: "skill" as const, name: "jyotish-vedic-astrology", status: "completed" as const, durationMs: 120 },
+      { sequence: 2, kind: "tool" as const, name: "run-jyotish-consultation", status: "completed" as const, durationMs: 1800 },
+    ],
+    stepBudget: { planned: 8, used: 2, remaining: 6, truncated: false },
+    workflow: workflowReceipt,
+    techniqueTruth: "verified",
+    techniqueAuditTable: [
+      { technique: "VedAstro Cloud State", status: "blocked" as const, note: "未经核验的官方云证据，置信度封顶" },
+      { technique: "Functional Benefic/Malefic", status: "executed" as const, note: "功能属性与自然属性冲突时降置信" },
+      { technique: "Formal Vargas D1–D60", status: "executed" as const, note: "20/20 传统命名分盘" },
+      { technique: "Vimshottari sub-periods", status: "executed" as const, note: "本轮已纳入证据计划" },
+      { technique: "Narayana Dasha", status: "executed" as const, note: "本轮无独立当前读数" },
+      { technique: "Yogas", status: "executed" as const, note: "成盘与落空均列出" },
+      { technique: "Ashtakavarga", status: "executed" as const, note: "本轮已纳入证据计划" },
+      { technique: "Shadbala components", status: "executed" as const, note: "本轮已纳入证据计划" },
+      { technique: "MEVG / Global Web Evidence", status: "blocked" as const, note: "本轮未闭合" },
+      { technique: "Real Case Calibration", status: "blocked" as const, note: "本轮未闭合" },
+      { technique: "Timing Precision Gate", status: "executed" as const, note: "按声明精度封顶" },
+      { technique: "Prashna chart", status: "not_applicable" as const, note: "本轮为本命咨询" },
+    ],
+  };
+  return {
+    techniqueTruth: "verified",
+    workflowReceipt,
+    agentExecutionReceipt,
+  };
+}
+
+function thinkingPlan(domains: readonly ("career" | "wealth" | "marriage")[]) {
+  return natalConsultationThinkingPlan({
+    domains,
+    requiredBlocks: [
+      "raw_structure",
+      "raman_six_step",
+      "yoga_table",
+      "timing",
+      "synthesis",
+      "technique_audit_table",
+      "modern_wrap",
+    ],
+    mustUseLayers: ["D1", "run-jyotish-consultation", "D10", "skill_read"],
+  });
+}
+
+function assistantMessage(_index: number, domains: readonly ("career" | "wealth" | "marriage")[]) {
+  const receipts = typicalReceipts(domains);
+  return {
+    role: "assistant" as const,
+    text: "正".repeat(4000),
+    thinkingText: "思".repeat(4000),
+    thinkingSections: thinkingPlan(domains),
+    ...receipts,
+  };
+}
+
+function userMessage(index: number) {
+  return {
+    role: "user" as const,
+    text: `第${String(index + 1).padStart(2, "0")}问`.padEnd(100, "问"),
+    requestId: `capacity-req-${index + 1}`,
+  };
+}
+
+function sessionDetailPayload(rounds: number, domains: readonly ("career" | "wealth" | "marriage")[]) {
+  const messages = Array.from({ length: rounds }, (_, index) => [
+    userMessage(index),
+    assistantMessage(index, domains),
+  ]).flat();
+  return {
+    session: {
+      id: "00000000-0000-4000-8000-000000000001",
+      title: "事业方向",
+      theme: "career",
+      model_id: "test-model",
+      messages,
+      session_type: "consultation",
+      rectification_case_id: null,
+      chart_profile_id: null,
+      chart_profile_name: null,
+      chart_profile_role: null,
+      updated_at: "2026-09-15T00:00:00.000Z",
+      pinned: false,
+      archived_at: null,
+    },
+  };
+}
+
+function countedText(messages: readonly { text?: string }[]): number {
+  return messages.reduce((sum, message) => sum + (message.text ?? "").length, 0);
+}
+
+function oldCountedChars(messages: readonly Record<string, unknown>[]): number {
+  return messages.reduce((sum, message) => {
+    const text = typeof message.text === "string" ? message.text.length : 0;
+    const thinking = typeof message.thinkingText === "string" ? message.thinkingText.length : 0;
+    const sections = "thinkingSections" in message
+      ? JSON.stringify(message.thinkingSections).length
+      : 0;
+    return sum + text + thinking + sections;
+  }, 0);
+}
+
+test("the capacity migration is the latest replace of append_consultation_question", () => {
+  const names = readdirSync(fileURLToPath(migrationsDir))
+    .filter((name) => name.endsWith(".sql"))
+    .sort();
+  const replaces = names.filter((name) =>
+    name.includes("append_consultation_question") || name.includes("consultation_session_capacity"),
+  );
+  assert.deepEqual(replaces.at(-1), "20260916010000_consultation_session_capacity.sql");
+  assert.equal(
+    existsSync(fileURLToPath(new URL("../db/migrations/20260916010000_consultation_session_capacity.sql", import.meta.url))),
+    false,
+    "business migration must not be copied into frontend/db/migrations (BUG-127/BUG-144)",
+  );
+});
+
+test("CREATE OR REPLACE keeps the BUG-464 signature, lock, idempotency, and error codes", () => {
+  assert.match(
+    sql,
+    /create or replace function public\.append_consultation_question\(\s*p_user_id uuid,\s*p_request_id text,\s*p_session_id uuid,\s*p_question_message jsonb\s*\)/,
+  );
+  assert.match(sql, /returns table\(success boolean, error_code text\)/);
+  const body = functionBody(sql);
+  assert.match(body, /pg_advisory_xact_lock/);
+  assert.match(body, /elem->>'requestId' = v_request_id/);
+  assert.match(body, /return query select true, null::text;/);
+  assert.match(body, /'session_missing'::text/);
+  assert.match(body, /'session_full'::text/);
+  assert.match(body, /'invalid_request'::text/);
+  assert.match(body, /'invalid_question_message'::text/);
+  assert.match(body, /char_length\(v_text\) > 16000/);
+  assert.match(body, /v_count >= 200/);
+  assert.match(body, /\(v_chars \+ v_new_chars\) > 200000/);
+  assert.match(
+    sql,
+    /revoke all on function public\.append_consultation_question\(uuid, text, uuid, jsonb\)\s+from public, anon, authenticated/,
+  );
+  assert.match(
+    sql,
+    /grant execute on function public\.append_consultation_question\(uuid, text, uuid, jsonb\)\s+to service_role/,
+  );
+  assert.doesNotMatch(body, /session_physically_full|quota_exceeded|storage_full/);
+});
+
+test("conversation quota sum counts only text; physical cap counts whole JSON", () => {
+  assert.match(originalSql, /length\(coalesce\(elem->>'thinkingText', ''\)\)/);
+  assert.match(originalSql, /elem \? 'thinkingSections'/);
+
+  const body = functionBody(sql);
+  const code = withoutLineComments(body);
+  assert.doesNotMatch(code, /thinkingText/);
+  assert.doesNotMatch(code, /thinkingSections/);
+  assert.match(code, /coalesce\(sum\(length\(coalesce\(elem->>'text', ''\)\)\), 0\)/);
+  assert.match(code, /coalesce\(sum\(length\(elem::text\)\), 0\)/);
+  assert.match(code, /\(v_physical_chars \+ v_new_physical\) > 1000000/);
+  assert.match(sql, /50 rounds × \(~4,000 body \+ ~4,000 thinkingText \+ ~3,000/);
+  assert.match(sql, /thinkingSections \+ ~3,000 receipts\) ≈ 700,000\. Headroom → 1,000,000\./);
+});
+
+test("both caps still surface as session_full; consult route is unchanged", () => {
+  const body = functionBody(sql);
+  const fullReturns = body.match(/return query select false, 'session_full'::text;/g) ?? [];
+  assert.equal(fullReturns.length, 1);
+  assert.match(consultRoute, /error_code === "session_full"/);
+  assert.match(consultRoute, /code: "session_full"/);
+  assert.doesNotMatch(consultRoute, /session_physically_full|physical_cap|storage_full/);
+  assert.match(itemRoute, new RegExp(`sessionSelect = "${SESSION_SELECT}"`));
+});
+
+test("measured thinkingSections and receipts stay inside the physical-cap arithmetic", () => {
+  const one = JSON.stringify(thinkingPlan(["career"]));
+  const two = JSON.stringify(thinkingPlan(["career", "wealth"]));
+  const three = JSON.stringify(thinkingPlan(["career", "wealth", "marriage"]));
+  assert.ok(one.length >= 1200 && one.length <= 2500, `1-domain sections JSON length ${one.length}`);
+  assert.ok(two.length >= 1800 && two.length <= 3500, `2-domain sections JSON length ${two.length}`);
+  assert.ok(three.length >= 2400 && three.length <= 4500, `3-domain sections JSON length ${three.length}`);
+
+  const receipts = typicalReceipts(["career", "wealth", "marriage"]);
+  const receiptJson = JSON.stringify({
+    techniqueTruth: receipts.techniqueTruth,
+    workflowReceipt: receipts.workflowReceipt,
+    agentExecutionReceipt: receipts.agentExecutionReceipt,
+  });
+  const receiptChars = receiptJson.length;
+  assert.ok(
+    receiptChars <= 4000,
+    `typical three-receipt JSON is ${receiptChars} chars; update the 3,000 term and 1,000,000 cap if this is a real stored shape`,
+  );
+
+  const perRound =
+    4000
+    + 4000
+    + three.length
+    + receiptChars;
+  const fiftyRoundPhysical = 50 * perRound;
+  assert.ok(
+    fiftyRoundPhysical < 1_000_000,
+    `50 × typical stored round (${perRound} chars) = ${fiftyRoundPhysical}; must stay under the physical cap`,
+  );
+});
+
+test("session-detail JSON at the old ~19-round cap versus the new ~50-round cap", () => {
+  const domains = ["career", "wealth", "marriage"] as const;
+  const oldCap = sessionDetailPayload(19, domains);
+  const newCap = sessionDetailPayload(50, domains);
+  const oldJson = JSON.stringify(oldCap);
+  const newJson = JSON.stringify(newCap);
+  const oldBytes = utf8Bytes(oldJson);
+  const newBytes = utf8Bytes(newJson);
+
+  assert.equal(oldCap.session.messages.length, 38);
+  assert.equal(newCap.session.messages.length, 100);
+  assert.ok(oldCountedChars(oldCap.session.messages) > 200_000, "19 full rounds already exceeded the old combined quota");
+  assert.ok(countedText(newCap.session.messages) >= 200_000, "50 rounds of 100+4000 visible text reach the conversation quota");
+  // Measured 2026-09-16 on this fixture: 19-round 562,240 B (0.536 MiB);
+  // 50-round 1,479,034 B (1.411 MiB); ratio 2.631.
+  assert.ok(oldBytes > 540_000 && oldBytes < 590_000, `19-round detail JSON was ${oldBytes} bytes`);
+  assert.ok(newBytes > 1_450_000 && newBytes < 1_510_000, `50-round detail JSON was ${newBytes} bytes`);
+  assert.ok(newBytes > oldBytes, "50-round detail JSON must be larger than 19-round");
+
+  const ratio = newBytes / oldBytes;
+  assert.ok(
+    ratio > 2.5 && ratio < 2.8,
+    `50/19 size ratio was ${ratio.toFixed(3)} (utf8 ${oldBytes} → ${newBytes})`,
+  );
+});

frontend/tests/database-consultation-session-capacity.test.ts

@@ -0,0 +1,277 @@
+import assert from "node:assert/strict";
+import { spawnSync } from "node:child_process";
+import { fileURLToPath } from "node:url";
+import test from "node:test";
+
+import { startPostgresFixture } from "./helpers/postgres-fixture.ts";
+
+const runnerPath = fileURLToPath(
+  new URL("../scripts/db-migrate.mjs", import.meta.url),
+);
+
+function dockerAvailable(): boolean {
+  return spawnSync("docker", ["version", "--format", "{{.Server.Version}}"], {
+    encoding: "utf8",
+    stdio: "ignore",
+  }).status === 0;
+}
+
+const skipWithoutDocker = dockerAvailable() ? false : "docker unavailable on this host";
+
+const SESSION_ID = "99999999-9999-4999-8999-999999999999";
+
+function sqlLiteral(value: string): string {
+  return value.replaceAll("'", "''");
+}
+
+function appendQuestionSql(
+  fixture: ReturnType<typeof startPostgresFixture>,
+  userId: string,
+  requestId: string,
+  textSql: string,
+  sessionId = SESSION_ID,
+): string {
+  return fixture.psql(`
+    select coalesce(success::text, 'null') || ':' || coalesce(error_code, 'null')
+    from public.append_consultation_question(
+      '${userId}'::uuid,
+      '${sqlLiteral(requestId)}',
+      '${sessionId}'::uuid,
+      jsonb_build_object('role', 'user', 'text', ${textSql})
+    );
+  `);
+}
+
+function appendQuestion(
+  fixture: ReturnType<typeof startPostgresFixture>,
+  userId: string,
+  requestId: string,
+  text: string,
+  sessionId = SESSION_ID,
+): string {
+  return appendQuestionSql(fixture, userId, requestId, `'${sqlLiteral(text)}'`, sessionId);
+}
+
+function seedSession(
+  fixture: ReturnType<typeof startPostgresFixture>,
+  userId: string,
+  messagesSql: string,
+): void {
+  fixture.psql(`
+    insert into public.chat_sessions (
+      id, user_id, title, theme, model_id, messages, session_type, updated_at
+    ) values (
+      '${SESSION_ID}', '${userId}', '新对话', 'general',
+      'test-model', ${messagesSql}, 'consultation', now()
+    )
+    on conflict (id) do update
+      set messages = excluded.messages,
+          title = excluded.title,
+          updated_at = now();
+  `);
+}
+
+test("append_consultation_question ignores thinking fields and enforces the physical JSON cap", { skip: skipWithoutDocker }, () => {
+  const fixture = startPostgresFixture();
+  const schemaUrl = fixture.connectionUrl("schema_owner", "schema-owner-test-password");
+
+  try {
+    const migration = spawnSync(process.execPath, [runnerPath], {
+      encoding: "utf8",
+      env: { ...process.env, SCHEMA_DATABASE_URL: schemaUrl },
+    });
+    assert.equal(migration.status, 0, migration.stderr);
+    assert.match(migration.stdout, /applied 20260916010000_consultation_session_capacity\.sql/);
+
+    fixture.psqlAs(
+      "identity_runtime",
+      "identity-runtime-test-password",
+      `
+        insert into identity.users (name, email, email_verified, email_verified_at)
+        values ('Capacity User', 'capacity@example.com', true, now());
+      `,
+    );
+    const userId = fixture.psql(
+      "select id from identity.users where email = 'capacity@example.com'",
+    );
+
+    assert.equal(
+      fixture.psql(`
+        select
+          has_function_privilege(
+            'service_role',
+            'public.append_consultation_question(uuid, text, uuid, jsonb)',
+            'execute'
+          ) || ':' ||
+          has_function_privilege(
+            'authenticated',
+            'public.append_consultation_question(uuid, text, uuid, jsonb)',
+            'execute'
+          ) || ':' ||
+          has_function_privilege(
+            'anon',
+            'public.append_consultation_question(uuid, text, uuid, jsonb)',
+            'execute'
+          )
+      `),
+      "true:f:f",
+    );
+
+    assert.equal(
+      appendQuestion(fixture, userId, "missing-session", "会话不存在"),
+      "false:session_missing",
+    );
+
+    seedSession(fixture, userId, "'[]'::jsonb");
+    assert.equal(appendQuestion(fixture, userId, "append-request-1", "第一问会不会丢"), "true:null");
+    assert.equal(appendQuestion(fixture, userId, "append-request-1", "不该写入的重复提问"), "true:null");
+    assert.equal(
+      fixture.psql(`select jsonb_array_length(messages) from public.chat_sessions where id = '${SESSION_ID}'`),
+      "1",
+    );
+
+    assert.equal(
+      appendQuestionSql(fixture, userId, "too-long", "repeat('x', 16001)"),
+      "false:invalid_question_message",
+    );
+
+    seedSession(fixture, userId, `
+      (
+        select coalesce(jsonb_agg(elem order by n), '[]'::jsonb)
+        from (
+          select n,
+            case when n % 2 = 1 then
+              jsonb_build_object(
+                'role', 'user',
+                'text', '短问题',
+                'requestId', 'think-seed-' || n::text
+              )
+            else
+              jsonb_build_object(
+                'role', 'assistant',
+                'text', '短回答',
+                'thinkingText', repeat('T', 4000),
+                'thinkingSections', jsonb_build_object('pad', repeat('S', 3000)),
+                'techniqueTruth', 'verified',
+                'workflowReceipt', jsonb_build_object(
+                  'route', 'career',
+                  'status', 'ready',
+                  'preciseTiming', 'allowed',
+                  'missingLayers', jsonb_build_array()
+                )
+              )
+            end as elem
+          from generate_series(1, 60) as n
+        ) as seeded
+      )
+    `);
+
+    const thinkingSums = fixture.psql(`
+      select
+        coalesce(sum(length(coalesce(elem->>'text', ''))), 0)::text
+        || ':' ||
+        coalesce(sum(
+          length(coalesce(elem->>'text', ''))
+          + length(coalesce(elem->>'thinkingText', ''))
+          + case when elem ? 'thinkingSections'
+              then length((elem->'thinkingSections')::text) else 0 end
+        ), 0)::text
+        || ':' ||
+        coalesce(sum(length(elem::text)), 0)::text
+      from public.chat_sessions,
+        jsonb_array_elements(coalesce(messages, '[]'::jsonb)) as elem
+      where id = '${SESSION_ID}'
+    `);
+    const [textOnly, oldFormula, physical] = thinkingSums.split(":").map(Number);
+    assert.ok(textOnly < 50_000, `visible text was ${textOnly}`);
+    assert.ok(oldFormula > 200_000, `old combined formula was ${oldFormula}`);
+    assert.ok(physical < 1_000_000, `physical JSON was ${physical}`);
+    assert.equal(appendQuestion(fixture, userId, "after-thinking", "思考不该占额度"), "true:null");
+    assert.equal(
+      fixture.psql(`select jsonb_array_length(messages) from public.chat_sessions where id = '${SESSION_ID}'`),
+      "61",
+    );
+
+    seedSession(fixture, userId, `
+      (
+        select coalesce(jsonb_agg(
+          jsonb_build_object('role', 'assistant', 'text', repeat('x', 19900), 'requestId', n::text)
+          order by n
+        ), '[]'::jsonb)
+        from generate_series(1, 10) as n
+      )
+    `);
+    const quotaBoundary = fixture.psql(`
+      select
+        coalesce(sum(length(coalesce(elem->>'text', ''))), 0)::text
+        || ':' ||
+        coalesce(sum(length(elem::text)), 0)::text
+      from public.chat_sessions,
+        jsonb_array_elements(coalesce(messages, '[]'::jsonb)) as elem
+      where id = '${SESSION_ID}'
+    `);
+    const [quotaText, quotaPhysical] = quotaBoundary.split(":").map(Number);
+    assert.equal(quotaText, 199_000);
+    assert.ok(quotaPhysical < 1_000_000, `quota-boundary physical JSON was ${quotaPhysical}`);
+    assert.equal(
+      appendQuestionSql(fixture, userId, "quota-full", "repeat('y', 1001)"),
+      "false:session_full",
+    );
+    assert.equal(
+      fixture.psql(`select jsonb_array_length(messages) from public.chat_sessions where id = '${SESSION_ID}'`),
+      "10",
+    );
+
+    seedSession(fixture, userId, `
+      jsonb_build_array(
+        jsonb_build_object(
+          'role', 'assistant',
+          'text', '短',
+          'techniqueTruth', 'verified',
+          'workflowReceipt', jsonb_build_object('route', 'career', 'status', 'ready', 'preciseTiming', 'allowed', 'missingLayers', jsonb_build_array()),
+          'agentExecutionReceipt', jsonb_build_object('pad', repeat('R', 1000001))
+        )
+      )
+    `);
+    const physicalOnly = fixture.psql(`
+      select
+        coalesce(sum(length(coalesce(elem->>'text', ''))), 0)::text
+        || ':' ||
+        coalesce(sum(length(elem::text)), 0)::text
+      from public.chat_sessions,
+        jsonb_array_elements(coalesce(messages, '[]'::jsonb)) as elem
+      where id = '${SESSION_ID}'
+    `);
+    const [shortText, hugePhysical] = physicalOnly.split(":").map(Number);
+    assert.ok(shortText < 100, `short body was ${shortText}`);
+    assert.ok(hugePhysical > 1_000_000, `receipt pad physical JSON was ${hugePhysical}`);
+    assert.equal(
+      appendQuestion(fixture, userId, "physical-full", "正文很短但行已经胀了"),
+      "false:session_full",
+    );
+    assert.equal(
+      fixture.psql(`select jsonb_array_length(messages) from public.chat_sessions where id = '${SESSION_ID}'`),
+      "1",
+    );
+
+    seedSession(fixture, userId, `
+      (
+        select coalesce(jsonb_agg(
+          jsonb_build_object('role', 'user', 'text', 'x', 'requestId', n::text)
+          order by n
+        ), '[]'::jsonb)
+        from generate_series(1, 200) as n
+      )
+    `);
+    assert.equal(
+      appendQuestion(fixture, userId, "count-full", "满了就不能再写"),
+      "false:session_full",
+    );
+    assert.equal(
+      fixture.psql(`select jsonb_array_length(messages) from public.chat_sessions where id = '${SESSION_ID}'`),
+      "200",
+    );
+  } finally {
+    fixture.stop();
+  }
+});

frontend/tests/database-local-business.test.ts

@@ -94,6 +94,7 @@ test("local PostgreSQL applies the reviewed business schema and serves authentic
     assert.match(migration.stdout, /applied 20260901020000_chat_session_pin_archive\.sql/);
     assert.match(migration.stdout, /applied 20260905010000_personal_report_longform_appendices\.sql/);
     assert.match(migration.stdout, /applied 20260915010000_rectification_touch_chat_session\.sql/);
+    assert.match(migration.stdout, /applied 20260916010000_consultation_session_capacity\.sql/);
     assert.equal(
       existsSync(fileURLToPath(new URL("../db/migrations/20260901020000_chat_session_pin_archive.sql", import.meta.url))),
       false,
@@ -104,6 +105,11 @@ test("local PostgreSQL applies the reviewed business schema and serves authentic
       false,
       "business migration must not be copied into frontend/db/migrations (BUG-127/BUG-144)",
     );
+    assert.equal(
+      existsSync(fileURLToPath(new URL("../db/migrations/20260916010000_consultation_session_capacity.sql", import.meta.url))),
+      false,
+      "business migration must not be copied into frontend/db/migrations (BUG-127/BUG-144)",
+    );
     fixture.psql(pinArchiveMigration);
     assert.equal(
       fixture.psql(`

tests/test_rectification_engine_memoization.py

@@ -2,17 +2,25 @@
 
 Golden payload in tests/golden/rectification_engine_memoization_v1.json was
 produced from origin/staging @ a8d29d1b before any memoization landed.
+
+Do not compare that payload with a whole-structure ``==``. Cross-machine
+libm / pyswisseph rounding already drifted ``margin_percent`` by 1.1e-3
+(TASK-rectification-engine-memoization-fix-20260915). Equivalence of the
+four cached layers is proven in-process instead.
 """
 
 from __future__ import annotations
 
 import inspect
 import json
+import math
 from datetime import date, datetime
 from pathlib import Path
 from typing import Any
 from unittest.mock import patch
 
+import pytest
+
 from scripts.active_rectification_event_engine import (
     _candidate_datetimes,
     _controlled_transit_rules,
@@ -35,6 +43,20 @@ GOLDEN_PATH = ROOT / "tests" / "golden" / "rectification_engine_memoization_v1.j
 FROZEN_TODAY = date(2026, 9, 16)
 TIMING_KEYS = frozenset({"column_compare_ms"})
 SOURCE_COMMIT = "a8d29d1b6cc37ff865ddec6c8bccdf9aa889ee53"
+CACHE_LAYER_KEYS = (
+    "ashtakavarga_result",
+    "shadbala_result",
+    "vimshottari_timeline",
+    "narayana_periods",
+)
+# TASK-rectification-engine-memoization-fix-20260915 §2 measured max
+# cross-machine drift of 1.1e-3 on decision_receipt.margin_percent
+# (5.2995 vs golden 5.3006). Candidate scores, already rounded to 4
+# decimals, drifted by 1 ulp (1.0e-4). Tolerance is the wider of
+# abs=2e-3 and rel=5e-4 so 1.1e-3 still passes while a 1e-2 mutation
+# (the reverse test below) fails.
+GOLDEN_FLOAT_ABS = 2e-3
+GOLDEN_FLOAT_REL = 5e-4
 
 
 def public_score_request() -> dict[str, Any]:
@@ -117,11 +139,102 @@ def _count_calls(monkeypatch, owner: Any, name: str, *, from_engine: bool = Fals
     return counter
 
 
+def _is_json_number(value: Any) -> bool:
+    return isinstance(value, (int, float)) and not isinstance(value, bool)
+
+
+def _assert_tiered_equal(actual: Any, expected: Any, *, path: str) -> None:
+    if isinstance(expected, dict):
+        assert isinstance(actual, dict), path
+        assert set(actual) == set(expected), f"{path} keys {set(actual)!r} != {set(expected)!r}"
+        for key in expected:
+            _assert_tiered_equal(actual[key], expected[key], path=f"{path}.{key}")
+        return
+    if isinstance(expected, list):
+        assert isinstance(actual, list), path
+        assert len(actual) == len(expected), f"{path} length {len(actual)} != {len(expected)}"
+        for index, (left, right) in enumerate(zip(actual, expected)):
+            _assert_tiered_equal(left, right, path=f"{path}[{index}]")
+        return
+    if isinstance(expected, int) and isinstance(actual, int) and not isinstance(expected, bool) and not isinstance(actual, bool):
+        assert actual == expected, f"{path}: {actual!r} != {expected!r}"
+        return
+    if _is_json_number(expected) or _is_json_number(actual):
+        left = float(actual)
+        right = float(expected)
+        tolerance = max(GOLDEN_FLOAT_ABS, GOLDEN_FLOAT_REL * abs(right))
+        assert math.isclose(left, right, rel_tol=0.0, abs_tol=tolerance), (
+            f"{path}: {left!r} vs {right!r} exceeds {tolerance}"
+        )
+        return
+    assert actual == expected, f"{path}: {actual!r} != {expected!r}"
+
+
+def _assert_memoization_payloads(actual: dict[str, Any], expected: dict[str, Any]) -> None:
+    _assert_tiered_equal(actual["candidate_scores"], expected["candidate_scores"], path="candidate_scores")
+    _assert_tiered_equal(actual["decision_receipt"], expected["decision_receipt"], path="decision_receipt")
+
+
+def _event_engine_request() -> dict[str, Any]:
+    """Dated events for compute_event_candidate_rows (not the v5 score payload)."""
+    return {
+        "birth_date": "1990-01-01",
+        "start_time": "12:00",
+        "end_time": "12:02",
+        "lat": 39.9,
+        "lon": 116.4,
+        "tz": 8,
+        "events": [
+            {
+                "id": "00000000-0000-4000-8000-000000000001",
+                "domain": "education",
+                "event_kind": "education_start",
+                "date": "2008-07-01",
+                "precision": "day",
+                "summary": "入学",
+            },
+            {
+                "id": "00000000-0000-4000-8000-000000000002",
+                "domain": "career",
+                "event_kind": "career_entry",
+                "date": "2012-06-15",
+                "precision": "day",
+                "summary": "入职",
+            },
+        ],
+    }
+
+
+def _contexts_with_layer_cache_cleared(contexts: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    cleared: list[dict[str, Any]] = []
+    for context in contexts:
+        clone = dict(context)
+        for key in CACHE_LAYER_KEYS:
+            clone[key] = None
+        cleared.append(clone)
+    return cleared
+
+
 def test_score_candidates_matches_baseline_golden() -> None:
     expected = json.loads(GOLDEN_PATH.read_text(encoding="utf-8"))
     actual = json.loads(json.dumps(_golden_payload(), ensure_ascii=True))
-    assert actual["candidate_scores"] == expected["candidate_scores"]
-    assert actual["decision_receipt"] == expected["decision_receipt"]
+    _assert_memoization_payloads(actual, expected)
+
+
+def test_golden_float_shift_of_1e_minus_2_fails() -> None:
+    expected = json.loads(GOLDEN_PATH.read_text(encoding="utf-8"))
+    mutated = json.loads(json.dumps(expected))
+    mutated["candidate_scores"][0]["score"] += 1e-2
+    with pytest.raises(AssertionError):
+        _assert_memoization_payloads(expected, mutated)
+
+
+def test_golden_discrete_field_mismatch_fails() -> None:
+    expected = json.loads(GOLDEN_PATH.read_text(encoding="utf-8"))
+    mutated = json.loads(json.dumps(expected))
+    mutated["candidate_scores"][0]["time"] = "99:99"
+    with pytest.raises(AssertionError):
+        _assert_memoization_payloads(expected, mutated)
 
 
 def test_shadbala_verified_fields_match_with_and_without_birth_minute() -> None:
@@ -372,3 +485,51 @@ def test_compute_event_candidate_rows_reuses_static_context_without_mutating_it(
     assert len(rows) == 2
     assert [frozenset(context) for context in contexts] == original_keys
     assert all("_transit_chart_cache" not in context for context in contexts)
+
+
+def test_cached_static_context_matches_uncached_fallback(monkeypatch) -> None:
+    request = _event_engine_request()
+    cached = [
+        build_candidate_static_context(request, candidate)
+        for candidate in _candidate_datetimes(request)
+    ]
+    assert len(cached) >= 2
+    assert len(request["events"]) >= 2
+    for context in cached:
+        for key in CACHE_LAYER_KEYS:
+            assert context.get(key) is not None
+    uncached = _contexts_with_layer_cache_cleared(cached)
+    for context in uncached:
+        for key in CACHE_LAYER_KEYS:
+            assert context[key] is None
+    calls = _count_calls(monkeypatch, event_engine.shadbala, "calc_shadbala")
+    rows_cached = compute_event_candidate_rows(request, static_contexts=cached)
+    cached_calls = calls[0]
+    rows_uncached = compute_event_candidate_rows(request, static_contexts=uncached)
+    uncached_calls = calls[0] - cached_calls
+    assert rows_cached == rows_uncached
+    assert cached_calls == 0
+    assert uncached_calls > cached_calls
+    assert uncached_calls >= len(cached) * len(request["events"])
+
+
+def test_poisoned_static_cache_diverges_from_live_rows() -> None:
+    request = _event_engine_request()
+    cached = [
+        build_candidate_static_context(request, candidate)
+        for candidate in _candidate_datetimes(request)
+    ]
+    rows_cached = compute_event_candidate_rows(request, static_contexts=cached)
+    poisoned = [dict(context) for context in cached]
+    poisoned[0]["shadbala_result"] = {
+        "planets": {
+            name: {
+                "sthana_bala": {"total": 0.0},
+                "drik_bala": 0.0,
+                "naisargika_bala": 0.0,
+            }
+            for name in ("Sun", "Moon", "Mars", "Mercury", "Jupiter", "Venus", "Saturn")
+        }
+    }
+    rows_poisoned = compute_event_candidate_rows(request, static_contexts=poisoned)
+    assert rows_poisoned != rows_cached