[loopx-agent]

Python-to-TypeScript journal writes can turn 1.0 into 1, lose signed zero, and round unsafe integers in saved evidence. The same numeric collisions can make a changed immutable plan look like a replay.

This change keeps the existing journal object carrier and reuses the sparse number metadata already used by context delivery. The Python collector now lives in the existing runtime IO adapter; context projection and the journal share one TypeScript validator. The journal owner restores the numeric spellings for publication and uses them in replay and immutable-plan comparisons. Goal lifetime, admission, effect, phase and lock checks remain with that owner. Ordinary object callers retain their bytes and operation identity; historical receipts are not rewritten.

The whole-JSON-string approach from the closed #7657 is superseded. It preserved numbers but double-escaped evidence text. On the same synthetic workload:

Quoted body	Existing object request	Whole-string interim	Sparse replacement
8,192 characters	11,411 B; numeric facts lost	14,651 B	11,663 B
65,536 characters	81,988 B; numeric facts lost	107,283 B	82,240 B
1,550,000 characters	1,909,020 B; numeric facts lost	2,505,262 B; rejected	1,909,272 B; accepted

The hard 2 MiB request limit and one-RPC write remain unchanged. The sparse arm preserves all original numeric facts and quoted text, and replay leaves the file unchanged. A number-dense control still grows from 30,270 B to 199,185 B; number-dense sources can reach the same request limit earlier. The incorrect baseline is not an efficiency target. These component checks do not establish a latency or long-horizon benefit.

Validation on final 56f971bed39290d9e0d63f8ce4b9b38ef8fa2eb8, based on main 2f544f969236f5dbbe858720108c41d07be60a6c:

• 86 integration cases pass; 15 optional native checkpoint profiles are not qualified. Actual File/SQLite journal, source capture/context and main checkpoint fixture paths are exercised.
• 182 TypeScript cases and 45 minimum-Node 22.22.3 cases pass.
• The rebuilt wheel matches all 1,898 delivered product files. Its 24 installed journal/CLI recovery cases pass; parent and Python helper imports stay in site-packages. All eight published full-capture readers use cat -- and are checked against the actual disposable capture references, including the final missing-evidence refusals.
• TypeScript typecheck, Ruff, configured mypy (17 targets), explicit IO-owner mypy (2 targets), bilingual RFC and documentation governance checks pass. Broader optional mypy still reports the same two pre-existing envelope typing errors on independent immutable base/head runs; both remain failed observations with typed attribution.
• Repository premerge passes:6 direct checks and14 selected checks, with the original120s per-check timeout;0 failures.

Earlier component aafc2d29 passed 203 source cases, 134 quiet CLI cases and 241 context cases (one Linux-only skip). All 11 proposed files are byte-identical in the final main integration; the final package/integration results above qualify the new combination separately. The initial concurrent CLI cohort's 131 passes and three startup/deadline failures are retained. The quiet rerun kept the original time limits. An initial package-audit wrapper also misclassified OS capture readers as Python readers; product tests passed, and the corrected audit follows the actual published commands.

This is an independent prerequisite for #7430. Unified composition, writer/consumer cutover, old-reader retirement, and comparable original-quota outcomes and full costs remain open there. The paired controller RFC and Turn protocol record this boundary and the rejected transport result.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; runtime_reported; reasoning_effort=max

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

Exact head: 56f971bed39290d9e0d63f8ce4b9b38ef8fa2eb8；不可变 base：2f544f969236f5dbbe858720108c41d07be60a6c。

动机 / Motivation

需要在中断后依据原始证据继续工作的 Agent，以及审阅 Turn journal 的维护者。

同一份计划含 1.0、-0.0 或非安全整数时，旧写入会改变数值；不同计划还可能被当作同一次重放。当前写入保留这些区别，并在改动不可变计划时拒绝覆盖。

同一真实写入探针在 base 复现四类失真和四种漏拒绝，当前源码与普通 wheel 均通过；普通对象的文件字节及 operation id 不变。

本切片不交付 C1 controller/consumer 整体切换、模型采用、长程净收益、合并或安装升级。

统一生产者、同 capture 消费与完整成本资格仍由既有 C1／#7430 路径继续验证。

改动思路 / Approach

原始 Python 对象继续作为 journal 的主体；只有 JavaScript number 不能区分的来源数值附带稀疏 token。TypeScript writer 在原子写入前恢复这些数值，同一来源区别还进入 operation id 和不可变 plan 的比较。只修落盘、却继续用已舍入的数字判断重放，不能修复这个问题。

复用已有 collector、数值校验器和 journal writer，比新增第二种 journal carrier 或 Python 决策 owner 更小。Python 仍只负责 IO 元数据；TypeScript 保留状态、阶段、来源准入与写入决定。该 PR 交付完整的数值保真前置，能够独立回退，不把 C1 全部迁移作为这一修复的前置条件。相邻重构已将 collector 和 validator 收敛为共享实现，没有保留两份数值判定。

具体改动 / Changes

完整审阅 11 个路径，+333/-72，包括两个共享数值适配文件、两个 context consumer、Python/TypeScript journal writer、两个测试文件及三份协议/RFC 文档。

关键代码讲解

• source_number_tokens 从现有 context adapter 移到共享 IO 边界。整值 float、带符号零和非安全整数带来源 token；普通安全整数、分数 double 沿原对象路径，非有限 float 在请求前拒绝。
• write_turn_journal 是真实 Python caller，只在有 token 时增加元数据，保留原对象、expected effect 和 source admission；仍调用同一个 turn_journal.write。
• sourceNumberTokens 校验路径类型、存在性、重复项、JSON 数值语法及解码值一致性。Context source identity 和 journal 共用这一实现，未引入文字或字段名猜测。
• commitTurnJournal 结合 parseJournal 保留当前及既存文件的数值区别。来源、effect、阶段和锁校验保持原 owner；拒绝不能覆盖原文件，恢复审计的浅拷贝也保留顶层数值映射。无 metadata 的普通对象继续原编码。

规范依据：LoopX Turn journal 原契约、C1 原边界。spec_revision 为上述完整 base。journal_source_fidelity、immutable_plan_replay、existing_authority_and_limits 在当前 head implemented；C1_cutover_and_net_benefit deferred。这里以修改前契约和真实反例作 oracle，本 PR 新增说明不证明自己的正确性。

对主干的风险 / Risks

未发现当前阻断性问题。最强风险是“文件中的数字看似正确，但重放仍把两个来源当作同一个计划”。同一公开 writer 探针在 base 丢失四类来源区别，四种计划改动均未被拒绝；当前 head 和普通 wheel 保留全部六组数值，并对四种改动报 plan is immutable，原文件字节不变。普通对象在 base/head/wheel 的字节与 operation id 完全一致，原有分数值及普通字符串也保留。

独立执行：176 项完整七模块 Python 测试、166 项相关 TypeScript 测试通过。覆盖实际 source admission、Goal 实例变化、阶段和 effect 冲突、重放、恢复、context delivery，以及非法 token／缺少 raw-number 支持的拒绝；当前 Node 为 24.21.0，未把它当作最低版本 Node 的新执行。普通 wheel 在空临时环境安装，1,787 个 Python/TypeScript/JSON 源成员与当前源码相等；真实 File/SQLite Goal 均完成 source-admitted 写入、重放、数字改动拒绝及独立 CLI 只读检查，registry 和主状态未变。最初陈旧前端资产的打包拒绝保留，随后按正常构建流程通过，没有绕过来源校验。

成本保留负面结果。同一带引号的 155 万字符正文，合成参数从对象的 1,907,777 字节变为稀疏方案的 1,907,840 字节；整段 JSON 字符串为 2,503,942，超过既有 2 MiB 上限。真实稀疏 journal 以一次 RPC 接收并保留正文，超大请求仍拒绝且不建文件。6,000 个整值 float 的参数则从 30,055 增至 190,971 字节：这是来源保真的成本，未包装成延迟、token 或长程收益。

语义与 CI 对齐

复用 loopx_turn_journal_v0 和既有临时 source_number_tokens 传输，不新增持久状态、权限或 actor lifecycle。两个 RFC 语言版本同改；双语 diff、TypeScript typecheck、diff-scoped advisory 和完整语义检查通过。标准 premerge 第一轮的 6 项直接检查与 14 项选择检查全过，其中语义扫描为 94.835 秒；补齐精确 scope 质量回执后复跑，同一 head 的语义扫描在原 120 秒超时。独立 immutable base 也在该上限超时。当前必需验证与 CQR 已如实保留失败，合并资格仍 hold；此前通过没有被删除，也不据此放行。

评审期间的脚本路径、外部脚本导入来源和不支持的 inspection registry 路由错误均保留；修正后以实际 module 路径、普通 wheel 与独立源数据回读核验。没有将未执行检查记为通过。按当前配置未查询、轮询或等待 CI。

我的整体评价 / Verdict

APPROVE，仅针对该完整精确 head 的数值保真修复，delivery 判断为 justified_increment。long_horizon 为 preserved：后续恢复所依赖的原始事实、不可变计划和一次结算身份得到保留；user_experience 为 improved：原命令和 journal 读回即可看见正确数字，不增加设置或确认回合。视觉审阅不适用：受影响的是 journal 与 CLI/context 传输，没有网页或 App 渲染改动；已安装 App、最低 Node、真实模型采用和长程收益仍未独立认证。

共享 collector/validator 和既有 writer 的边界与问题相称；没有必要增加新 codec、配置或捕获服务。C1 的完整切换和收益验收仍开放，代码批准不授予 merge、dismiss、安装或新的实验权限。

English verdict: APPROVE — 56f971b. Matched real-writer and installed-wheel probes preserve numeric source facts, immutable-plan refusal and ordinary-object identity; 176 Python and 166 TS tests pass. C1 cutover and end-to-end benefit remain separate; the current required semantic-deadline and CQR gates remain held.