Read a short subject's only number as its reading

A two-character Chinese subject — 温度, 室温, 设定, 预算 — is the common
way to state a reading, and the rename rule was swallowing its number:
「温度 22 度」 became a thing called 温度22 with nothing left to report, so
the lane recorded no conflict at all and no later memory could disagree
with it. A bare number written onto the subject is a name only while the
memory still has another number waiting behind it; with nothing after it,
it is the reading. Measured across 33 pairs: 6 change — five missed
conflicts now recorded, plus one new disclosed boundary (「Room 3 is big」
and 「Room 4 is big」 read as one quantity that moved, because a label
cannot tell an identifier from a reading). Everything else is byte-equal.

Also corrected: a memory that is only a number does have a subject — the
number itself — so 「22 度」 and 「24 度」 record nothing because they name
two different things, not because the store found nothing to compare.

The smoke script now refuses an answer from a server that predates the
memory tiers: it requires recall_count in the remember response, a field
the pre-p2-3 Go build never emits.
This commit is contained in:
ZacharyZhang-NY committed 2026-10-09 11:44:29 -04:00
1 parent 8bed319848
commit 11e31eb7e3
4 files changed
+73 -14

No files matched your search

+5 -6
View File
@@ -331,7 +331,7 @@
- [x] p1-6.3 删 v1/v2/v3 废弃端口 - [x] p1-6.3 删 v1/v2/v3 废弃端口
- 量完的结论是**没有可删的**,不是没做:每个 `v1/v2/v3` 都承重——磁盘格式标签(`ely-eye-memory-capsule-v1` `proofs/checks_cartridge.py:128`、`ely-eye-{text,visual}-vectors-v1`、`ely-eye-kivi-replay-v1` `cartridge_assets.py:100,127,158`)、迁移阶梯(`migrations.py:181,186,239,241,243`,行号从 commit 本身量:`git show <sha>:path | grep -n`,工作区会被 pi-lens 重排而漂 9 行)、矛盾语料 fixture 名(`proofs/contradictions.py:114-135`)、第三方名字(MiniLM-L12-v2、ScreenSpot-v2、LongBench v2)、OpenAI 兼容的 `/v1` URL、缓存键里的 `compiler-v2` salt。没有废弃端口、别名或双代码路径 - 量完的结论是**没有可删的**,不是没做:每个 `v1/v2/v3` 都承重——磁盘格式标签(`ely-eye-memory-capsule-v1` `proofs/checks_cartridge.py:128`、`ely-eye-{text,visual}-vectors-v1`、`ely-eye-kivi-replay-v1` `cartridge_assets.py:100,127,158`)、迁移阶梯(`migrations.py:181,186,239,241,243`,行号从 commit 本身量:`git show <sha>:path | grep -n`,工作区会被 pi-lens 重排而漂 9 行)、矛盾语料 fixture 名(`proofs/contradictions.py:114-135`)、第三方名字(MiniLM-L12-v2、ScreenSpot-v2、LongBench v2)、OpenAI 兼容的 `/v1` URL、缓存键里的 `compiler-v2` salt。没有废弃端口、别名或双代码路径
- [x] p1-6.4 注释压到 ≤1 行 - [x] p1-6.4 注释压到 ≤1 行
- 口径:`tokenize` 找连续 `#` 行块。全仓 **39 块 / 86 行 → 0 块**,每块压成一行,压的时候留承重那句,实测数字与错误原文不删(`Unable to allocate 582. MiB for an array with shape (5632, 9024, 3)`、`288,266 rows for 165,973 atoms`、`[^[^\W\s]` 在 3.14 会吃掉 CJK) - 口径:`tokenize` 找连续的**整行** `#` 块(行尾注释不算)。全仓 **39 块 / 86 行 → 0 块**,每块压成一行,压的时候留承重那句,实测数字与错误原文不删(`Unable to allocate 582. MiB for an array with shape (5632, 9024, 3)`、`288,266 rows for 165,973 atoms`、`[^[^\W\s]` 在 3.14 会吃掉 CJK)
- docstring 不动(全仓多行 docstring 114 条,与父提交同数):docstring 是模块与函数的契约,不是注释;p1-4 review 用的口径也是「新增多行 docstring 数不得增加」。第一轮 review 抓到搬家途中塌了 3 条——`store_search.py` 的 `search_atoms`(那条「每个过滤条件都是同一条字面 SQL 里的 `? IS NULL OR` 谓词」规则和 `as_holder` 的共享规则)和 `FtsPage`、`bench/run.py` 的模块 docstring——已按原文补回,行数仍全部在 500 以内 - docstring 不动(全仓多行 docstring 114 条,与父提交同数):docstring 是模块与函数的契约,不是注释;p1-4 review 用的口径也是「新增多行 docstring 数不得增加」。第一轮 review 抓到搬家途中塌了 3 条——`store_search.py` 的 `search_atoms`(那条「每个过滤条件都是同一条字面 SQL 里的 `? IS NULL OR` 谓词」规则和 `as_holder` 的共享规则)和 `FtsPage`、`bench/run.py` 的模块 docstring——已按原文补回,行数仍全部在 500 以内
- 门禁:`pytest -q` **269 passed**(那是 `9cb46c1` 当时的套件规模,p2-1 加完测试已到 287);`ruff check backend/src/ely_eye backend/tests bench` All checks passed;mypy 冷缓存(`.venv/bin/python -m mypy --cache-dir <新目录> backend/src bench`)**24 errors / 9 files**,与父提交同命令量到的 **24 / 6** 净新增 0——同样 24 条错,`training.py` 拆开后落到 4 个新文件里,所以文件数从 6 变 9;`pyright` 查该 commit 改动的 **46 个文件**(`git show --name-only 9cb46c1` = 46 个 `.py` + 1 个 `.md`,**0 个前端文件**)→ **4 errors**,与 HEAD 单个 `training.py` 在包副本里量到的 4 条同源(qwen_vl_utils 元组返回、两处 `save_pretrained(Path)`、`bitsandbytes` 未解析);go build/vet/test + gofmt 干净;`npm run build` + eslint 干净 - 门禁:`pytest -q` **269 passed**(那是 `9cb46c1` 当时的套件规模,p2-1 加完测试已到 287);`ruff check backend/src/ely_eye backend/tests bench` All checks passed;mypy 冷缓存(`.venv/bin/python -m mypy --cache-dir <新目录> backend/src bench`)**24 errors / 9 files**,与父提交同命令量到的 **24 / 6** 净新增 0——同样 24 条错,`training.py` 拆开后落到 4 个新文件里,所以文件数从 6 变 9;`pyright` 查该 commit 改动的 **46 个文件**(`git show --name-only 9cb46c1` = 46 个 `.py` + 1 个 `.md`,**0 个前端文件**)→ **4 errors**,与 HEAD 单个 `training.py` 在包副本里量到的 4 条同源(qwen_vl_utils 元组返回、两处 `save_pretrained(Path)`、`bitsandbytes` 未解析);go build/vet/test + gofmt 干净;`npm run build` + eslint 干净
- herdr review(同一个 `p1review` pi,`strata/qwen3.8-flash-next-abliterated-iq3_s:high`)三轮:**round 1 FAIL 4** → **round 2 FAIL 1** → **round 3 PASS**。round 1 它自己跑出门禁、行数、搬体同一性、死代码、注释压缩、v1/v2/v3 全部承重,报的 4 条里 1 条是真产品缺陷:拆文件时掉了一些多行 docstring(全仓 114→111),掉的是契约不是装饰——`search_atoms` 的「每个过滤条件都是同一条字面 SQL 里的 `? IS NULL OR` 谓词」规则和 `as_holder` 共享规则、`FtsPage` 的游标理由、`bench/run.py` 的五臂说明(commit `28959c5` 按原文补回,114 条同父提交);其余 3 条是我写错的数:`store_search.py`/`bench/run.py`/`query_set.py` 行数、`cli.py` 调用点数(实为 11 个调用点亗 7 个方法)。round 2 剩 1 条:`migrations.py` 行号引成了 248/250,在 commit 里那是 PRAGMA 和空行——**根因是我在工作区量的,而 pi-lens 会在本轮之外重排文件把行号推下 9 行**;从此行号一律 `git show <sha>:path | grep -n` 从 commit 量(commit `9682657`)。round 3 它逐条重验:迁移阶梯 181/186/239/241/243 全中(并指出 `:243` 本身不带版本 token,`grep "v1\|v2"` 拓不到),CHECKLIST 那一行其余引用全部重验(`checks_cartridge.py:128`、`cartridge_assets.py:100/127/158`、`contradictions.py:114/135`、`config.py:20/27`、`cache_fabric.py:21`、`benchmarks.py:85`、`bench/build_screenspot.py:1/28`),`git diff --stat 28959c5..9682657` 只动一行,门禁重跑 269 passed / ruff / filesize_check 全绿 - herdr review(同一个 `p1review` pi,`strata/qwen3.8-flash-next-abliterated-iq3_s:high`)三轮:**round 1 FAIL 4** → **round 2 FAIL 1** → **round 3 PASS**。round 1 它自己跑出门禁、行数、搬体同一性、死代码、注释压缩、v1/v2/v3 全部承重,报的 4 条里 1 条是真产品缺陷:拆文件时掉了一些多行 docstring(全仓 114→111),掉的是契约不是装饰——`search_atoms` 的「每个过滤条件都是同一条字面 SQL 里的 `? IS NULL OR` 谓词」规则和 `as_holder` 共享规则、`FtsPage` 的游标理由、`bench/run.py` 的五臂说明(commit `28959c5` 按原文补回,114 条同父提交);其余 3 条是我写错的数:`store_search.py`/`bench/run.py`/`query_set.py` 行数、`cli.py` 调用点数(实为 11 个调用点亗 7 个方法)。round 2 剩 1 条:`migrations.py` 行号引成了 248/250,在 commit 里那是 PRAGMA 和空行——**根因是我在工作区量的,而 pi-lens 会在本轮之外重排文件把行号推下 9 行**;从此行号一律 `git show <sha>:path | grep -n` 从 commit 量(commit `9682657`)。round 3 它逐条重验:迁移阶梯 181/186/239/241/243 全中(并指出 `:243` 本身不带版本 token,`grep "v1\|v2"` 拓不到),CHECKLIST 那一行其余引用全部重验(`checks_cartridge.py:128`、`cartridge_assets.py:100/127/158`、`contradictions.py:114/135`、`config.py:20/27`、`cache_fabric.py:21`、`benchmarks.py:85`、`bench/build_screenspot.py:1/28`),`git diff --stat 28959c5..9682657` 只动一行,门禁重跑 269 passed / ruff / filesize_check 全绿
@@ -409,16 +409,15 @@
- [x] p2-3.1 工作记忆层:`working <holder> <session>` 按记录顺序返回该 session 记下的记忆,不经检索;HTTP `GET /api/memory/working` 与 MCP `working_set` 走同一份实现(Go 只转发,写规则仍在一处)。空 session 直接拒:`a working set needs a session; without one every memory is working memory`。验收:`pytest backend/tests/test_memory_tiers.py backend/tests/test_memory_cli.py backend/tests/test_memory_api.py`(13 + 14 + 8,合跑 35 passed)+ `bash scripts/mcp_write_smoke.sh` 里跨进程那一段(session 名不合并两个 holder;陌生 session 返回空而不是别人的记忆)。提交 `dc9cfb5`。 - [x] p2-3.1 工作记忆层:`working <holder> <session>` 按记录顺序返回该 session 记下的记忆,不经检索;HTTP `GET /api/memory/working` 与 MCP `working_set` 走同一份实现(Go 只转发,写规则仍在一处)。空 session 直接拒:`a working set needs a session; without one every memory is working memory`。验收:`pytest backend/tests/test_memory_tiers.py backend/tests/test_memory_cli.py backend/tests/test_memory_api.py`(13 + 14 + 8,合跑 35 passed)+ `bash scripts/mcp_write_smoke.sh` 里跨进程那一段(session 名不合并两个 holder;陌生 session 返回空而不是别人的记忆)。提交 `dc9cfb5`。
- [x] p2-3.2 巩固台账:`consolidations` 表按 `(holder, scenario)` 存 watermark + rollup atom id,一条 lane 只有一份常驻摘要、原地替换;`lane_atoms` 用严格 `ingested_at > watermark`,所以 rollup 自己不会被再算一次。**实测出一个真缺陷**:`watermark_at` 原来写成本轮时钟,用落后于记忆的时钟重跑会把同一批算两遍(第二次 `atom_count=4`),现改成 `max(atom.ingested_at for atom in batch)`。验收:`test_a_lane_rolls_up_once_and_keeps_rolling_from_where_it_stopped`(第二次返回 None,第三次 `atom_count=3` 且 `rollup_atom_id` 不变)。 - [x] p2-3.2 巩固台账:`consolidations` 表按 `(holder, scenario)` 存 watermark + rollup atom id,一条 lane 只有一份常驻摘要、原地替换;`lane_atoms` 用严格 `ingested_at > watermark`,所以 rollup 自己不会被再算一次。**实测出一个真缺陷**:`watermark_at` 原来写成本轮时钟,用落后于记忆的时钟重跑会把同一批算两遍(第二次 `atom_count=4`),现改成 `max(atom.ingested_at for atom in batch)`。验收:`test_a_lane_rolls_up_once_and_keeps_rolling_from_where_it_stopped`(第二次返回 None,第三次 `atom_count=3` 且 `rollup_atom_id` 不变)。
- [x] p2-3.3 时间/价值字段:schema v10 给 atom 加 `session_id` / `importance` / `recall_count` / `last_recalled_at` / `archived_at`;`v_atoms_current` 加 `archived_at IS NULL`,且 `apply()` 先 `DROP VIEW v_atoms_current`——`CREATE VIEW IF NOT EXISTS` 会留着旧视图体、把归档当没看见,升级测试钉住这点。**`decay` 没有落成字段**:没测过的衰减公式就是编故事,所以归档规则把条件写在命令里(`--older-than-days`、`--max-importance`、`recall_count = 0`),`importance` 是人声明的、`recall_count` 是 store 自己观察的。`importance` **未设也算不值钱**(SQL 是 `importance IS NULL OR importance <= ?`),rule 字符串早先只写 `importance at or under 0.5`、漏了 NULL,被评审判为 D30,现写 `importance unset or at or under 0.5`。验收:`test_a_search_counts_the_memory_it_handed_back`、`test_a_memory_a_search_already_returned_is_not_stale`、`test_an_archived_memory_stops_answering_without_leaving_the_store`、`test_the_archive_rule_states_its_own_terms_and_can_be_read_before_it_acts`(dry-run 先点名,`atom_count` 不减)、`test_a_restore_counts_the_memories_it_brought_back_and_only_those`(`restore_atoms` 加了 `archived_at IS NOT NULL`:一个从未归档的 id 早先打印 `{"restored": 1}` 却什么都没变)。 - [x] p2-3.3 时间/价值字段:schema v10 给 atom 加 `session_id` / `importance` / `recall_count` / `last_recalled_at` / `archived_at`;`v_atoms_current` 加 `archived_at IS NULL`,且 `apply()` 先 `DROP VIEW v_atoms_current`——`CREATE VIEW IF NOT EXISTS` 会留着旧视图体、把归档当没看见,升级测试钉住这点。**`decay` 没有落成字段**:没测过的衰减公式就是编故事,所以归档规则把条件写在命令里(`--older-than-days`、`--max-importance`、`recall_count = 0`),`importance` 是人声明的、`recall_count` 是 store 自己观察的。`importance` **未设也算不值钱**(SQL 是 `importance IS NULL OR importance <= ?`),rule 字符串早先只写 `importance at or under 0.5`、漏了 NULL,被评审判为 D30,现写 `importance unset or at or under 0.5`。验收:`test_a_search_counts_the_memory_it_handed_back`、`test_a_memory_a_search_already_returned_is_not_stale`、`test_an_archived_memory_stops_answering_without_leaving_the_store`、`test_the_archive_rule_states_its_own_terms_and_can_be_read_before_it_acts`(dry-run 先点名,`atom_count` 不减)、`test_a_restore_counts_the_memories_it_brought_back_and_only_those`(`restore_atoms` 加了 `archived_at IS NOT NULL`:一个从未归档的 id 早先打印 `{"restored": 1}` 却什么都没变)。
- [x] p2-3.4 冲突记录持久化:`conflicts` 表存(旧, 新, 共享标签, 两个值, 是否已裁决),只在同 holder 同 scenario 且**数字**变了时记录;纯措辞分歧不记(换句话就会每次触发)。数字按「主语 + 最近标签」双键归一(实测 `numeric_values('恒温器放在书房,设定 22 度')` 交出的值键是 `恒温` 与 `设定`,不是 `书房`——评审第 2 轮 D38 纠正的就是这里)。**冲突必须两侧主语相同**(`trust.quantity_conflict`),而且**写在主语上的裸数字是名字、不是量**(`_subject_and_rename_slot`)。评审第 1 轮实测出 D31:只共享谓语的不同事实早先被记成冲突(`客厅的恒温器调到 22 度` + `卧室的恒温器调到 24 度` → `调到` 22→24;`她买了 3 本书` + `他买了 5 本书` → `买了` 3→5),后果是 `compile("客厅的恒温器现在几度")` 的 `packed_text` 只剩卧室那条——问客厅拿到卧室的数字。第 2 轮又实测出 D37,方向相反:主语相同、但那个数字其实在给对象编号,也被当成一个量(`Room 3 holds 22 degrees` + `Room 4 holds 24 degrees` → `room` 3→4,搜 "degrees" 只剩 Room 4;`The router 1 forwards 220 packets` + `The router 2 forwards 240 packets` → `router` 1→2;`2026 saw 3 releases` + `2027 saw 5 releases` → `saw` 2026→2027)。现在的规则:主语之前独立成词的裸数字并入主语键、不当值(`room3` vs `room4`、`router1` vs `router2`、`2026` vs `2027` → 各 0 条),而**同一个主语**下数字变了仍是冲突(`The router forwards 220/240 packets` → 报 `router`,因为主语自己带着值;`2026 saw 3 releases` + `2026 saw 5 releases` → `saw` 3→5,早先读出 0 条,属漏报)。数字只要站在某个词之后就是量不是名字(`The rent is 2200 dollars` + `The rent is 2400 dollars` → `rent` 2200→2400;`恒温器 is 22 degrees` + `恒温器 is 24 degrees` → `恒温` 22→24),CJK 串里融合的数字也仍是量(`温度22度` + `温度24度` → `温度` 22→24)。主语落在虚词后面要扫到第一个实词(`The router …` 的主语是 `router` 不是 `the`),扫不到主语就不记(fail-closed)。明写的限制(每条都有测试钉住):store 比的是标签不是语义,同主语量不同东西仍算一个量——`恒温器设定 22 度` 与 `恒温器功率 100 瓦` 记一条冲突,由人裁决(`test_the_same_subject_measuring_two_things_still_reads_as_one_quantity`);给主语加限定词会被当成另一个东西,这是**漏报**(`恒温器调到 22 度` 然后 `客厅的恒温器调到 24 度` → 0 条,主语从 `恒温` 变成 `客厅`;两条都留着、什么都没藏,只是这次更新没被当成不一致,人仍可 `supersede`);只有**数字**能给主语改名,字母不行、英文限定词也不行,这是中英不对称的**误报**(`Room A holds 22 degrees` + `Room B holds 24 degrees` → 1 条;`The router model X forwards 220 packets` + `The router forwards 240 packets` → 1 条,而中文 `三楼的恒温器` / `四楼的恒温器` 会分开成 0 条);一条记忆只有数字没有词时没有主语可比(`22 度` + `24 度` → 0 条)。检索抑制有**同页条件**:`losing_sides` 要求旧新两侧都在候选 id 里,页面先排宽到 `limit * CANDIDATE_WINDOW`(=10)再裁(`_rank` 内部已截断),实测 `top_k=1` 仍给新侧;只有新侧完全没被召回时旧值才留在页上,此时读者看到的是被替换过的数字、且没有冲突标记。裁决后两侧都回来且 `valid_to` 仍为 None——裁决不等于 supersede。验收:`backend/tests/test_memory_conflicts.py`(22)+ MCP `memory_conflicts`(Go 4 条测试,`go test ./internal/mcpserver/` 按名跑 4 PASS)。 - [x] p2-3.4 冲突记录持久化:`conflicts` 表存(旧, 新, 共享标签, 两个值, 是否已裁决),只在同 holder 同 scenario 且**数字**变了时记录;纯措辞分歧不记(换句话就会每次触发)。数字按「主语 + 最近标签」双键归一(实测 `numeric_values('恒温器放在书房,设定 22 度')` 交出的值键是 `恒温` 与 `设定`,不是 `书房`——评审第 2 轮 D38 纠正的就是这里)。**冲突必须两侧主语相同**(`trust.quantity_conflict`),而且**写在主语上的裸数字是名字、不是量**(`_subject_and_rename_slot`)。评审第 1 轮实测出 D31:只共享谓语的不同事实早先被记成冲突(`客厅的恒温器调到 22 度` + `卧室的恒温器调到 24 度` → `调到` 22→24;`她买了 3 本书` + `他买了 5 本书` → `买了` 3→5),后果是 `compile("客厅的恒温器现在几度")` 的 `packed_text` 只剩卧室那条——问客厅拿到卧室的数字。第 2 轮又实测出 D37,方向相反:主语相同、但那个数字其实在给对象编号,也被当成一个量(`Room 3 holds 22 degrees` + `Room 4 holds 24 degrees` → `room` 3→4,搜 "degrees" 只剩 Room 4;`The router 1 forwards 220 packets` + `The router 2 forwards 240 packets` → `router` 1→2;`2026 saw 3 releases` + `2027 saw 5 releases` → `saw` 2026→2027)。现在的规则:主语之前独立成词的裸数字并入主语键、不当值(`room3` vs `room4`、`router1` vs `router2`、`2026` vs `2027` → 各 0 条),而**同一个主语**下数字变了仍是冲突(`The router forwards 220/240 packets` → 报 `router`,因为主语自己带着值;`2026 saw 3 releases` + `2026 saw 5 releases` → `saw` 3→5,早先读出 0 条,属漏报)。数字只要站在某个词之后就是量不是名字(`The rent is 2200 dollars` + `The rent is 2400 dollars` → `rent` 2200→2400;`恒温器 is 22 degrees` + `恒温器 is 24 degrees` → `恒温` 22→24),CJK 串里融合的数字也仍是量(`温度22度` + `温度24度` → `温度` 22→24)。**但写在主语上的裸数字只有后面还留着别的数字时才是名字**(`_numbers_follow`):句子里再没有第二个数字时,那个数字就是这条记忆要报的量——`温度 22 度` + `温度 24 度`、`室温`、`设定`、`预算 5000 元` + `预算 8000 元`、`The thermostat 22 degrees` + `The thermostat 24 degrees` 各 1 条(评审第 3 轮 D40:早先两字主语后的数字被并进主语、`values` 空,整条 lane 一条都不记,三字主语 `恒温器 22/24 度` 却记 1 条——主语长度决定记不记,当时没写)。主语落在虚词后面要扫到第一个实词(`The router …` 的主语是 `router` 不是 `the`),扫不到主语就不记(fail-closed)。明写的限制(每条都有测试钉住):store 比的是标签不是语义,同主语量不同东西仍算一个量——`恒温器设定 22 度` 与 `恒温器功率 100 瓦` 记一条冲突,由人裁决(`test_the_same_subject_measuring_two_things_still_reads_as_one_quantity`);给主语加限定词会被当成另一个东西,这是**漏报**(`恒温器调到 22 度` 然后 `客厅的恒温器调到 24 度` → 0 条,主语从 `恒温` 变成 `客厅`;两条都留着、什么都没藏,只是这次更新没被当成不一致,人仍可 `supersede`);只有**数字**能给主语改名,字母不行、英文限定词也不行,这是中英不对称的**误报**(`Room A holds 22 degrees` + `Room B holds 24 degrees` → 1 条;`The router model X forwards 220 packets` + `The router forwards 240 packets` → 1 条,而中文 `三楼的恒温器` / `四楼的恒温器` 会分开成 0 条);主语后面没有第二个数字时,编号会被当成量——`Room 3 is big` + `Room 4 is big` → 1 条(`room` 3→4),标签分不出编号和读数,这是误报方向、交人裁决(`test_an_identifier_with_no_reading_after_it_reads_as_a_quantity`);值键仍会把主语之后的第二个数字挂到最近的标签上(`Room 3 holds 22 degrees at 5 pm` 的 `degrees` 键拿到 `5`),本轮比对先看主语键所以没暴露后果(评审精度注 P5);两条都只剩数字的记忆各有各的主语(`numeric_values('22 度').subject` 是 `22`、`'24 度'` 是 `24`,主语不同所以 0 条——评审第 3 轮 D41 纠正的就是我原先写的「没有主语可比」)。检索抑制有**同页条件**:`losing_sides` 要求旧新两侧都在候选 id 里,页面先排宽到 `limit * CANDIDATE_WINDOW`(=10)再裁(`_rank` 内部已截断),实测 `top_k=1` 仍给新侧;只有新侧完全没被召回时旧值才留在页上,此时读者看到的是被替换过的数字、且没有冲突标记。裁决后两侧都回来且 `valid_to` 仍为 None——裁决不等于 supersede。验收:`backend/tests/test_memory_conflicts.py`(26)+ MCP `memory_conflicts`Line truncated
- 验收(原文两条): - 验收(原文两条):
- LongMemEval 风格两类**有分数**:`bench/reports/LEGS.md` 的 memory 表把 150 条可答 conversation 查询**整集**测了(不是 15 条抽样),shipped 配置下 knowledge-update **1.000**、temporal-reasoning **0.879**、multi-session 0.857、single-session-preference 0.800、single-session-user 0.960;开 dense 腿后 temporal-reasoning **0.950**、multi-session 0.920、preference 0.960。报告另有一列「最新的那条是否第一个回来」:words-only 0.52 / 0.48 / 0.37,dense 0.40 / 0.46 / 0.47——recall 满分不等于排序把被替换的旧值让开了位。 - LongMemEval 风格两类**有分数**:`bench/reports/LEGS.md` 的 memory 表把 150 条可答 conversation 查询**整集**测了(不是 15 条抽样),shipped 配置下 knowledge-update **1.000**、temporal-reasoning **0.879**、multi-session 0.857、single-session-preference 0.800、single-session-user 0.960;开 dense 腿后 temporal-reasoning **0.950**、multi-session 0.920、preference 0.960。报告另有一列「最新的那条是否第一个回来」:words-only 0.52 / 0.48 / 0.37,dense 0.40 / 0.46 / 0.47——recall 满分不等于排序把被替换的旧值让开了位。
- 被 supersede 的旧事实不当作现行:`v_atoms_current` 过滤 `valid_to`(`backend/tests/test_memory_schema.py:163 test_v_atoms_current_hides_closed_retired_and_unbegun_memories`),Go store 读的也是这个视图(p0-2 时实测跨语言断过),smoke 里 supersede 之后旧 atom 不再出现在 `search_atoms` 结果里。 - 被 supersede 的旧事实不当作现行:`v_atoms_current` 过滤 `valid_to`(`backend/tests/test_memory_schema.py:163 test_v_atoms_current_hides_closed_retired_and_unbegun_memories`),Go store 读的也是这个视图(p0-2 时实测跨语言断过),smoke 里 supersede 之后旧 atom 不再出现在 `search_atoms` 结果里。
- 分层没动检索质量:同一 harness、同一 store、同一 300 探针,p2-3 前后 LEGS 每个 recall 格子**逐字相同**(0.847 / 0.863 / 0.873 / 0.860 / 0.873 / 0.817 / 0.848,memory 表同样),因为这批语料里没有冲突行也没有归档行——新增的是状态与规则,不是排序器。复现法(评审 D33 要求可复算):`git worktree add /tmp/pre_p23 4c6af87` 只把产品退回 p2-3 之前、harness 保持当前、同一 store 跑 `python -m bench.leg_check`——我量过一次,独立评审在同一对照下也量过一次,格子逐字相同(stamp 显示 schema v9,因为 pre-v10 产品不会迁移);这条对照要在主仓里跑并 `PYTHONPATH=<worktree>/backend/src`——worktree 里没有 `bench/runs/*`,直接跑会死在 `123 store sources are not corpus files`(`bench/ingest_baseline.py:87` 的 `source_map()` 按 cwd 解析语料路径)。延迟只引提交产物里的数:`arms[1]`「words only (the configured 24)」臂 p50 **122.5 ms**(load [17.23,10.83,6.53],`bench/reports/leg_check.json`);评审重跑同一 harness 得到 91.2 ms 而每个 recall 格子逐字不变——延迟随机器负载走,报告页脚已写明跨 run 不可比(早先两次跑的 101.6 / 87.5 ms 已被后一次覆盖,不作证据)。 - 分层没动检索质量:同一 harness、同一 store、同一 300 探针,p2-3 前后 LEGS 每个 recall 格子**逐字相同**(0.847 / 0.863 / 0.873 / 0.860 / 0.873 / 0.817 / 0.848,memory 表同样),因为这批语料里没有冲突行也没有归档行——新增的是状态与规则,不是排序器。复现法(评审 D33 要求可复算):`git worktree add /tmp/pre_p23 4c6af87` 只把产品退回 p2-3 之前、harness 保持当前、同一 store 跑 `python -m bench.leg_check`——我量过一次,独立评审在同一对照下也量过一次,格子逐字相同(stamp 显示 schema v9,因为 pre-v10 产品不会迁移);这条对照要在主仓里跑并 `PYTHONPATH=<worktree>/backend/src`——worktree 里没有 `bench/runs/*`,直接跑会死在 `123 store sources are not corpus files`(`bench/ingest_baseline.py:87` 的 `source_map()` 按 cwd 解析语料路径)。延迟只引提交产物里的数:`arms[1]`「words only (the configured 24)」臂 p50 **122.5 ms**(load [17.23,10.83,6.53],`bench/reports/leg_check.json`);评审重跑同一 harness 得到 91.2 ms 而每个 recall 格子逐字不变——延迟随机器负载走,报告页脚已写明跨 run 不可比(早先两次跑的 101.6 / 87.5 ms 已被后一次覆盖,不作证据)。
- 门禁:`pytest -q` 379 passed;六文件 78 passed(tiers 13 / conflicts 22 / schema 12 / cli 14 / api 8 / product_api 9);ruff clean;mypy 全仓冷缓存 24 errors/9 files = `9cb46c1` 基线(净 0);pyright 触碰文件 0 errors;go build/vet/test + gofmt clean;`npm run build` 产物 `index-CSDHCFNo.js` 未变(本轮未动前端);`bench.audit --product` → `audited 32 harness files, 2 reports, 6 set files` / clean;`mcp_write_smoke.sh` PASS(主语规则上线后跨进程那条 24/22 冲突仍成立)。 - 门禁:`pytest -q` 383 passed;六文件 82 passed(tiers 13 / conflicts 26 / schema 12 / cli 14 / api 8 / product_api 9);三文件(tiers+cli+api)35 passed;ruff clean;`bench.filesize_check` clean;mypy 全仓冷缓存 24 errors/9 files = `9cb46c1` 基线(净 0);pyright 触碰文件 0 errors;go build/vet/test + gofmt clean;`npm run build` 产物 `index-CSDHCFNo.js` 未变(本轮未动前端);`bench.audit --product` → `audited 32 harness files, 2 reports, 6 set files` / clean;`mcp_write_smoke.sh` PASS(主语规则上线后跨进程那条 24/22 冲突仍成立)。
- **评审第 1 轮 FAIL**(同一独立 pi,8 项全跑):一条行为缺陷 + 一条规则串与行为不符 + 四条记录——D31 只共享谓语的不同事实被记成冲突(客厅/卧室、她/他,后果是问客厅拿到卧室的数字);D30 归档 rule 字符串漏了 `importance IS NULL` 也归档;D32 「只给新的一侧」少了同页条件;D33 「前后逐字相同」没有留档产物(改成可复算的对照路径);D34 引了没提交进产物的 p50;D35 引了不存在的测试名。评审同时实测确认:门禁全绿且 mypy 与 `9cb46c1` 基线逐字相同、pyright 0、go 全套、`index-CSDHCFNo.js` 未变、smoke PASS;删掉主语相等那三行的反向实验恰好让那三条新测试变红;它自己重跑 `bench.leg_check`,32 行表格里除两列延迟外全部与提交产物相同,并量到 scratch store 里冲突 0 条 / 巩固 0 条 / 归档 0 行(所以「这批语料没有冲突行也没有归档行」从断言变成实测)。
- **评审第 2 轮 FAIL**(同一独立 pi,8 项全跑,D30–D35 的修法全部实测成立):四条——D36 纪律:本轮新增两处整行 `#` 连续两行(`trust.py:234-235`、`:270-271`),并且它更正了自己第 1 轮的口径:按本仓 p1-6.4 那条规则(tokenize、任意缩进)`9cb46c1` 是 0 块、`deabbf7` 是 4 块(`memory.py:325`、`migrations.py:318`、`test_memory_schema.py:287`、`bench/leg_check.py:326`)、`3daface` 是 6 块,六处全部收成一行;D37 记录与行为不符且**方向与我写的相反**(见上面 p2-3.4 那条:误报同样会把一条记忆从答案页上摘掉,我原先只写了漏报一侧);D38 「最近标签是 `书房`」实测是 `设定`;D39 三处措辞错(「两个方呰」「然候」「主语自己变了就点主语」——代码条件是主语**自己带着**那个值)。修法不是加披露而是换机制:写在主语上的裸数字是名字不是量。评审另给两条精度说明(不算缺陷):那条 worktree 对照直接跑会死在 `123 store sources are not corpus files`,要在主仓里跑并 `PYTHONPATH=<worktree>/backend/src`;「shipped 臂」是我起的标签,产物里的名字是 `arms[1]`「words only (the configured 24)」(来自 `config.py:43 retrieval_top_k: int = 24`)。
- D36–D39 修完的门禁:`pytest -q` **379 passed**(conflicts 17→22,六文件 78、三文件 35)、ruff 干净、`bench.filesize_check` 干净、mypy 冷缓存 **24 errors / 9 files**(净 0)、pyright 对触碰文件 **0 errors**、go 全套 + `gofmt -l` 空、`npm run build` 产物未变、`bench.audit --product` 32 / 2 / 6 clean、`mcp_write_smoke.sh` PASS、全仓整行 `#` 连续 ≥2 的块 **0**(tokenize 扫全部 tracked `.py`)。
- **评审第 1 轮 FAIL**(同一独立 pi,8 项全跑,~2 h / 462 工具轮):D30 rule 字符串漏 NULL、D31 标签粒度把不同事实当冲突(真行为缺陷,见 p2-3.4)、D32 没写同页条件、D33 「前后逐字相同」无产物、D34 引了未提交的 p50、D35 引了不存在的测试名——六条已全部按实测改掉,并新增 8 条测试钉住边界(两间房、两个人、一台路由器、同主语不同量、无主语、主语加了限定词、首数字当值、restore 只数真正归档过的)。该轮同时实测确认:全部门禁数字与本报告一致、两个反向实验复现、归档 dry-run 零副作用(sqlite `mtime_ns` 前后相同)、22→24→26 链产生 3 条冲突、跨 holder/跨 scenario 各 0 条、`supersede` 不记冲突、裁决后两侧回且 `valid_to` 仍 None、harness blob 三处全同、json 与 markdown 四表逐字一致、它自己的 pre-p2-3 对照复现了每个 recall 格子。 - **评审第 1 轮 FAIL**(同一独立 pi,8 项全跑,~2 h / 462 工具轮):D30 rule 字符串漏 NULL、D31 标签粒度把不同事实当冲突(真行为缺陷,见 p2-3.4)、D32 没写同页条件、D33 「前后逐字相同」无产物、D34 引了未提交的 p50、D35 引了不存在的测试名——六条已全部按实测改掉,并新增 8 条测试钉住边界(两间房、两个人、一台路由器、同主语不同量、无主语、主语加了限定词、首数字当值、restore 只数真正归档过的)。该轮同时实测确认:全部门禁数字与本报告一致、两个反向实验复现、归档 dry-run 零副作用(sqlite `mtime_ns` 前后相同)、22→24→26 链产生 3 条冲突、跨 holder/跨 scenario 各 0 条、`supersede` 不记冲突、裁决后两侧回且 `valid_to` 仍 None、harness blob 三处全同、json 与 markdown 四表逐字一致、它自己的 pre-p2-3 对照复现了每个 recall 格子。
- **评审第 2 轮 FAIL**(同一独立 pi,8 项全跑,D30–D35 的修法全部实测成立):四条——D36 纪律:本轮新增两处整行 `#` 连续两行(`trust.py:234-235`、`:270-271`),并且它更正了自己第 1 轮的口径:按本仓 p1-6.4 那条规则(tokenize、任意缩进)`9cb46c1` 是 0 块、`deabbf7` 是 4 块(`memory.py:325`、`migrations.py:318`、`test_memory_schema.py:287`、`bench/leg_check.py:326`)、`3daface` 是 6 块,六处全部收成一行;D37 记录与行为不符且**方向与我写的相反**(见上面 p2-3.4 那条:误报同样会把一条记忆从答案页上摘掉,我原先只写了漏报一侧);D38 「最近标签是 `书房`」实测是 `设定`;D39 三处措辞错(「两个方呰」「然候」「主语自己变了就点主语」——代码条件是主语**自己带着**那个值)。修法不是加披露而是换机制:写在主语上的裸数字是名字不是量。评审另给两条精度说明(不算缺陷):那条 worktree 对照直接跑会死在 `123 store sources are not corpus files`,要在主仓里跑并 `PYTHONPATH=<worktree>/backend/src`;「shipped 臂」是我起的标签,产物里的名字是 `arms[1]`「words only (the configured 24)」(来自 `config.py:43 retrieval_top_k: int = 24`)。
- **评审第 3 轮 FAIL**(同一独立 pi,8 项全跑,~534 工具轮;D36–D39 的修法全部实测成立,16 对冲突逐对复现):两条——D40 机制盲区:主语是两字 CJK 词或单个拉丁词时,紧跟的裸数字被并进主语、`values` 空,`_note_conflicts` 直接返回,整条 lane 一条不记(`温度 22/24 度`、`室温`、`设定`、`预算 5000/8000 元`、`The thermostat 22/24 degrees` 全部 0 条,HTTP 上也是 `[]`),而三字主语 `恒温器 22/24 度` 记 1 条——**主语长度决定记不记**,且 `:412` 那句「数字只要站在某个词之后就是量不是名字」被 `温度 22 度` 直接推翻;D41 归因错:我写「一条记忆只有数字没有词时没有主语可比」,实测 `numeric_values('22 度').subject == '22'`,0 条是因为两侧主语不同。修法换成机制(`_numbers_follow`)而不是加披露,实测 33 对里只动 6 对(五条漏报变记上,加一条新披露边界 `Room 3 is big` + `Room 4 is big` → 1 条),其余逐字不变。该轮同时实测确认:门禁全绿、五个攻击(标点隔开、数字开头、融合写法、一句多数字、反向实验)全部量过,`Room 3 holds 22 degrees at 5 pm` 报 `room3` 22→24 是真读数变化、但 `degrees` 键拿到了 `5`(P5,本轮无害);反向实验删掉改名那四行恰好让两条测试变红(我预测的第三条不变红,因为「数字开头当主语」住在 `_subject_and_rename_slot` 第一个分支 `trust.py:254`);抑制只在页面(搜只给新侧,`GET /api/atoms` 两侧都在且 `valid_to` 仍 None)。另有四条精度说明:P1 smoke 要加指纹断言,否则 8799 上的旧服务能替这棵树答题(它自己就撞上一次:答题体没有 p2-3 字段)——已做:`scripts/mcp_write_smoke.sh:175` 现在要求 `remember` 的答复里有 `recall_count`(Go 侧无 `omitempty`,所以必发;`git show 4c6af87:ely-eye-mcp/internal/domain/types.go` 里 0 处、HEAD 1 处,旧构建缺这个字段就会当场 FAIL),这个断言确实会中断跑(字段名写错的那一版在正常跑里就报过 exit 1);P2 「整行」口径(已写进 p1-6.4);P3 本报告里门禁行与第 1 轮记录行各重复一份(已去重);P4 `LEGS.md` 的 stamp(`a52c104`,harness blob `a1e9842d`)比改名后的 HEAD(blob `977d19eb`)旧——实测 `git diff a52c104:bench/leg_check.py HEAD:bench/leg_check.py` 只有一行注释被压成一行、零代码变动,`bench/leg_report.py` 两处 blob 全同,所以报告里的数字就是这份代码跑出来的,不必重测(重测只会把延迟换成另一个机器负载下的值)。
- 顺带:`cli.py` 495→432 行(写路径拆进 `cli_memory.py`),`bench/leg_check.py` 拆出 `bench/leg_report.py`;typer 0.27.2 用**关键字**调用回调,`*atom_ids: str` 形式的命令一调就 `TypeError`,改成 `list[str]`——这才是之前 `test_maintenance` 那条 TypeError 的真因,命令重名只是表象。报告 stamp 早于 harness 的缺陷也修了(`a52c104`)。 - 顺带:`cli.py` 495→432 行(写路径拆进 `cli_memory.py`),`bench/leg_check.py` 拆出 `bench/leg_report.py`;typer 0.27.2 用**关键字**调用回调,`*atom_ids: str` 形式的命令一调就 `TypeError`,改成 `list[str]`——这才是之前 `test_maintenance` 那条 TypeError 的真因,命令重名只是表象。报告 stamp 早于 harness 的缺陷也修了(`a52c104`)。
- [ ] **p2-4** 共享记忆 - [ ] **p2-4** 共享记忆
- [ ] p2-4.1 多用户/多 agent 作用域 + ACL - [ ] p2-4.1 多用户/多 agent 作用域 + ACL
+21 -5
View File
@@ -225,7 +225,9 @@ def numeric_values(text: str) -> StatedNumbers:
differently but lead with the same thing, and it is that shared label which makes the pair one differently but lead with the same thing, and it is that shared label which makes the pair one
disagreement about one quantity rather than two unrelated counts. A number written onto the disagreement about one quantity rather than two unrelated counts. A number written onto the
subject itself is part of the object's name, not a quantity — 「Room 3」 and 「Room 4」 are two subject itself is part of the object's name, not a quantity — 「Room 3」 and 「Room 4」 are two
rooms — so it joins the subject instead of being recorded as a value. rooms — but only while the memory still has another number waiting to be its reading. A number
with nothing after it is the reading itself: 「温度 22 度」 states a temperature, and renaming the
thermostat after it would leave the memory with no quantity left to compare.
""" """
try: try:
tokens = fts_terms(text) tokens = fts_terms(text)
@@ -257,20 +259,34 @@ def _subject_and_rename_slot(tokens: list[str]) -> tuple[str, int]:
A memory that opens with a number is about that number: 「2026 saw 3 releases」 counts releases A memory that opens with a number is about that number: 「2026 saw 3 releases」 counts releases
for 2026, so 2026 names the thing and 3 measures it. An English memory usually opens with an for 2026, so 2026 names the thing and 3 measures it. An English memory usually opens with an
article, and a subject behind one is still the subject: 「The router forwards 220 packets」 leads article, and a subject behind one is still the subject: 「The router forwards 220 packets」 leads
with its router, not with 「the」. Only a number written onto the subject renames it; one with its router, not with 「the」. Only a number written onto the subject renames it, and only
standing past a word is a value, which is why 「The rent is 2200 dollars」 can move. A CJK run can when the memory still has a later number to report: 「Room 3 holds 22 degrees」 names a room and
then reads it, while 「温度 22 度」 has nothing behind the 22 but the unit. A number standing past
a word is a value either way, which is why 「The rent is 2200 dollars」 can move. A CJK run can
fuse the subject and what it reports into one token, and the number after that is a value too: fuse the subject and what it reports into one token, and the number after that is a value too:
「恒温器调到 24 度」 adjusts a thermostat, it does not name a thing called 「恒温器调到 24」. 「恒温器调到 24 度」 adjusts a thermostat, it does not name a thing called 「恒温器调到 24」.
""" """
for index, token in enumerate(tokens): for index, token in enumerate(tokens):
if _NUMBER.fullmatch(token.strip(_PUNCTUATION)): stated = _NUMBER.fullmatch(token.strip(_PUNCTUATION))
if stated:
return token.strip(_PUNCTUATION), -1 return token.strip(_PUNCTUATION), -1
head = _label(token, first=True) head = _label(token, first=True)
if head: if head:
return head, -1 if _label(token) != head else index + 1 renames = _label(token) == head and _numbers_follow(tokens, index + 2)
return head, index + 1 if renames else -1
return "", -1 return "", -1
def _numbers_follow(tokens: list[str], start: int) -> bool:
"""Whether the memory still has a bare number to state after the slot in question.
A name is only worth taking when something else is left to be the reading. Without this test the
shortest Chinese phrasings — 「温度 22 度」, 「预算 5000 元」 — fold their only number into the
subject and the memory reports no quantity at all, so no later memory can ever disagree with it.
"""
return any(_NUMBER.fullmatch(token.strip(_PUNCTUATION)) for token in tokens[start:])
def quantity_conflict(older_text: str, newer_text: str) -> QuantityConflict | None: def quantity_conflict(older_text: str, newer_text: str) -> QuantityConflict | None:
"""The quantity two memories state differently, but only while they lead with the same subject. """The quantity two memories state differently, but only while they lead with the same subject.
+42 -3
View File
@@ -198,12 +198,12 @@ def test_the_same_subject_measuring_two_things_still_reads_as_one_quantity(tmp_p
) )
def test_a_memory_that_leads_with_a_number_has_no_subject_to_conflict_on(tmp_path) -> None: def test_two_memories_that_are_only_a_number_do_not_name_the_same_thing(tmp_path) -> None:
service, _ = _lane(tmp_path, ["22 度", "24 度"]) service, _ = _lane(tmp_path, ["22 度", "24 度"])
assert service.conflicts("wife") == [], ( assert service.conflicts("wife") == [], (
"the number is all the memory has to name itself with, so the store has no subject it could " "each number becomes its own subject — 「22」 and 「24」 — so these are two different things "
"call the two one thing and records nothing" "to the store, not one quantity that moved; it finds a subject, it just is not the same one"
) )
@@ -279,3 +279,42 @@ def test_a_letter_qualifier_does_not_rename_the_room(tmp_path) -> None:
"room whose reading moved — the pair is recorded for the holder to settle rather than " "room whose reading moved — the pair is recorded for the holder to settle rather than "
"silently resolved, and this is the boundary the rule does not cross" "silently resolved, and this is the boundary the rule does not cross"
) )
def test_a_short_chinese_reading_is_a_value_and_not_part_of_a_name(tmp_path) -> None:
service, _ = _lane(tmp_path, ["温度 22 度", "温度 24 度"])
found = service.conflicts("wife")
assert [conflict.terms for conflict in found] == [["温度"]], (
"「温度 22 度」 has nothing behind the 22 but the unit, so the 22 is the reading and not the "
"name of a thing called 温度22 — renaming it would leave the memory stating no quantity at "
"all, and the most common Chinese phrasing would never be noticed as an update"
)
assert [(conflict.older_value, conflict.newer_value) for conflict in found] == [("22", "24")]
def test_a_budget_that_changed_is_recorded(tmp_path) -> None:
service, _ = _lane(tmp_path, ["预算 5000 元", "预算 8000 元"])
found = service.conflicts("wife")
assert [(conflict.terms, conflict.older_value, conflict.newer_value) for conflict in found] == [
(["预算"], "5000", "8000")
], "a two-character subject is the common case, not an edge case: the number after it moves"
def test_a_latin_reading_with_nothing_after_it_is_a_value_too(tmp_path) -> None:
service, _ = _lane(tmp_path, ["The thermostat 22 degrees", "The thermostat 24 degrees"])
assert len(service.conflicts("wife")) == 1, (
"the same rule has to hold in both scripts or it is a rule about Chinese, not about memories"
)
def test_an_identifier_with_no_reading_after_it_reads_as_a_quantity(tmp_path) -> None:
service, _ = _lane(tmp_path, ["Room 3 is big", "Room 4 is big"])
assert len(service.conflicts("wife")) == 1, (
"a label cannot tell an identifier from a reading: with no second number in the sentence the "
"3 and 4 are compared as one quantity that moved, so this pair is recorded and the holder "
"settles it — the boundary the rule does not cross, and the price of not hiding a disagreement"
)
+5
View File
@@ -172,6 +172,11 @@ for required in ("remember", "supersede_memory", "forget_memory", "mounted_cartr
statement = "keys are in the hall drawer" statement = "keys are in the hall drawer"
remembered = tool("remember", {"text": statement, "scenario": "home"})["atom"] remembered = tool("remember", {"text": statement, "scenario": "home"})["atom"]
if "recall_count" not in remembered:
raise SystemExit(
"FAIL: the API answered with no recall_count; a server left over from before the memory tiers "
"took the port after our guard passed — stop it and re-run"
)
if not remembered["holder_id"]: if not remembered["holder_id"]:
raise SystemExit(f"FAIL: the server stamped no holder: {remembered}") raise SystemExit(f"FAIL: the server stamped no holder: {remembered}")
atom_id = remembered["atom_id"] atom_id = remembered["atom_id"]