feat(progress-review): close the drift loop with typed receipts and a recorded differential

[songoow]

What this PR does

Closes the progress-review loop through existing LoopX contracts and ships a reproducible differential. Revised on 2026-09-22 after an external review of the first closed-loop version; the four defects it found are fixed deterministically (table below), and the matrix was re-recorded with question set scoped-progress-sentinel-v2.

• Core capability progress-review-sentinel (default off). Per-goal policy control_plane.progress_review with off / shadow / assist, a drift signal, a drift_threshold, and a pinned contract_revision. A strict progress_review_receipt_v0 contract whose drift booleans the core recomputes from the typed judgments and rejects when they disagree. A pure trigger joins receipts to run rows by turn identity (with Agent/Todo agreement), counts consecutive completed drift receipts bound to the pinned revision, skips at most two newest still-pending evaluations, and in assist raises the existing autonomous_replan_obligation after the typed repeat fuse. The refresh-state writeback judges an acknowledgement against the same obligation, so an accepted replan re-arms it. The policy vocabulary lives in the control plane; the capability re-exports it, so the control plane never imports the capability layer.
• Policy surfaces. configure-goal --progress-review-mode|--progress-review-signal|--progress-review-drift-threshold|--progress-review-contract-revision|--clear-progress-review-configuration, the configuration catalog, the Dashboard capability editor (assets repackaged), the chat configuration API and localization.
• Observer (packages/loopx-jev). The model receives only operator basis fields plus scoped material, never goal identity (byte-identical requests across identities are pinned by a test). Three Noul questions ask about the change between checkpoints, not the after state; drift init --runtime-root writes a pending receipt on queue and the evaluated receipt on drain; drift label records private truth; status reports observer_influence and the fact that consumption is governed by the Goal policy.
• Differential harness and closed-loop test. loopx-jev sentinel compare replays a frozen 16-sequence matrix against committed recordings; test_closed_loop.py runs one real refresh-state sequence through off, shadow, unpinned assist (blocked, reason reported), pinned assist (obligation raised, shown in loopx status, acknowledged, re-armed).

Review findings and fixes

Finding (2026-09-21 review)	Fix in this revision
Harness passed a goal_id derived from the case name into model input	Model state is a whitelist of basis fields; hashed goal ids in the harness; anonymity test
noul gated on behaviour change: unrelated features passed, negative experiments were flagged	Rule progress_review_signal_rule_v1 gates on serves_acceptance and evidence_increment, both asked about the marginal change; core recomputes and rejects inconsistent receipts
Receipts only had to agree with each other, not with the current goal contract	assist requires a pinned contract_revision; other revisions are stale and never counted; unpinned assist reports contract_revision_unpinned
Turn-id join ignored Agent/Todo; an unevaluated newest run dissolved the streak	Identity agreement required; ambiguous fallbacks unattributed; pending receipts skipped within a bound of two
worker_influence: none misdescribed assist	Observer reports observer_influence + core_consumption; docs state that assist changes the work contract (required obligation) while granting no new authority

CI failures on the previous head are addressed: mypy Callable type, the control-plane import boundary, the ratchet oversized function (helper extracted), the packaged dashboard assets, and a rebase onto 0d464d2ac that picks up upstream fixes for the goal-channel and SQLite Node tests.

Recorded differential (committed v2 live recording, jev-1.13.0, 35 rounds, 34 completed, 1 failed closed)

	Typed repeat fuse	noul receipts	choice receipts
Drift sequences flagged at their gold round	0/9 (invisible by construction)	9/9	5/9
Reaching the assist obligation at threshold 2	0/9	9/9	2/9
Real on-goal upstream commits falsely flagged	0/7	0/7	0/7
Premature flags inside mixed sequences	0	0	0

Cosmetic sequences flag at round 1; mixed sequences flag at exactly their drift round (serves_acceptance 0.06–0.13 on post-implementation churn, versus 0.62–0.90 under the v1 wording). The negative probe (evidence_increment 0.85) and the necessary failing test (serves_acceptance 0.73) stay unflagged. A second independent live run reproduced every outcome on all 16 sequences. Client latency 1.45–1.49 s median and 2.9 s p95 in the v2 runs (0.74–0.81 s median in earlier runs; network and provider time), 1890 median input tokens.

Caveats. The v2 wording was revised after seeing the v1 misses on these constructed sequences, so they are not held-out evidence for the wording; the seven real commits were not used to tune anything. Constructed gold labels come from their author. One of 35 answers was rejected by the strict decoder (selected label ≠ argmax) and fails closed. No claim of reduced wasted work or task-completion improvement: that needs a held-out set labelled with drift label and a separately authorized intervention study.

Verification

• 242 tests across the package (real Git, real refresh-state, receipts, labels, pending lifecycle, anonymity, matrix strictness, replay determinism, closed loop) and the touched core suites (trigger, capability, configuration UI, chat API, registry, refresh gate, contract quality, architecture boundary, maintainability ratchet).
• ruff check on the CI scope, kernel mypy (22 files), strict mypy on the new modules, docs-governance-smoke, repository-hygiene-smoke, progress-review-sentinel-smoke, packaged dashboard rebuild.

Draft exit and merge readiness

This PR asks for stage 0 of the readiness ladder in loopx/capabilities/progress_review/README.md: a default-off registered capability whose shadow records receipts and whose assist requires a pinned contract revision. Author-side conditions, all met on this head:

• CI green on the exact head (26 pass, 5 conditional skips)
• External review findings fixed deterministically, not by tuning the model
• Differential replays from committed recordings without a credential; second live run reproduced it
• Bilingual capability README, operation guide and decision record, including an adoption path and the readiness ladder
• No new authority: no pause, gate, settlement or acceptance; the core never calls a model or reads a raw delta

Merging a control-plane change is the maintainers' decision and is never self-merged. Stages 1 (shadow with labels on real Goals) and 2 (one pinned assist Goal) are operator studies made with the tool after merge; stage 3 (defaults, escalation, pause) needs a separately authorized intervention study and is out of scope.

Boundaries

No pause, user gate, Turn settlement or Goal acceptance authority; escalation legs are future work. The core never calls a model, never reads a raw delta, never overwrites progress_observation.

[huangruiteng]

Draft design review：把进展评估接入同一条 replan 闭环

针对 head 6be153723eaefddc3d251a7a2ddf5eacfa4cf534。本次是围绕“与现有 replan 融合、改进触发语义并伴随 TS 重构”的设计反馈和定向代码验证，不是全量 merge-readiness 认证。PR 当前仍为 Draft；loopx pr-review 为该 head 返回 no-action，fresh-audit 又要求已有有效结论，因此本次不伪称完成工具未生成的正式评审计划，也不发布批准。

动机

我支持研究这个问题：Agent 可以持续报告 advanced，而实际工作没有推进当前目标。精确重复检测确实不能覆盖所有这种情形。但更有价值的交付应当是增强现有 replan 的证据和触发质量，而不是让一个外部模型形成第二套规划权威。

这与总路线的长程纠偏、可恢复执行有关，也可以成为 TS 规则收敛的合理切片。参考 overall roadmap 和 TS migration RFC 的 Delivery semantics。原 semantic-assistance RFC 的 M0 讨论稿收录不能代替 assist 的产品采用决定；本评论提出整合方向，不认证 Jev 的模型质量或默认启用资格。

改动思路

目前 PR 已做对一件重要的事：external_progress_review_drift 最终进入已有 autonomous_replan_obligation，status 与 refresh writeback 也使用同一个 receipt loader。这是应该保留的方向。

建议进一步收敛为：

运行历史 / Todo·Vision 事实 / 可选外部评估 receipt
        ↓ 绑定 Goal·Agent·工作对象·当前基准，校验时效和去重
领域触发器 → 类型化触发事实 → 明确的优先级与触发策略
        ↓
现有 autonomous_replan_obligation / action packet
        ↓
现有 replan_semantics.ts 校验实际 writeback → ACK / 后续重触发

模型负责回答有限的语义问题；确定性规则决定何时要求复盘；原有 Agent、Todo/Vision 与 writeback owner 决定如何处理。外部评估不能取消已有硬义务，模型不可用也不能使原有 replan 失效。复盘允许有证据地继续原计划，不应强迫每次制造新 Todo 或改方向。

具体改动

本 head 的范围包括：内置 capability、Goal 配置和 CLI/API/Dashboard 编辑器；receipt codec/存储/投影；run-history trigger 与 refresh-state ACK 接入；可选 observer 的 capture/queue/drain/provider 调用；对照 harness、合成与真实提交 fixture、双语说明。以下意见主要针对 assist 闭环，其余表面尚未得到全量合并资格验证。

1. [P1] 先修复“原样重复也能结清漂移义务”

obligation builder 的 review 分支 只携带 review 信息，没有像 typed_progress_repeat 一样带入 progress_baseline；trigger 本身也没有提供该 baseline。

我在隔离 runtime 中经过真实 refresh_state_run 和 receipt 写入复现：两轮各有 surface/hypothesis/evidence 的 advanced 记录，被两份完成的 drift receipt 触发 assist 后，原样提交最新一轮的同一 surface、hypothesis、evidence，仍得到 semantic_delta.accepted=true、new_surface/new_hypothesis、durable ACK，随后 obligation 消失。把同一 observation 作为 baseline 调用已有 novelty evaluator，则返回 accepted=false。

原因是 semantic_delta_from_writeback 找不到 baseline，semantic_progress_delta 就把已有标识算作新增。这不是模型准确率问题，而是当前闭环可以被零增量写回解除。

最小修复：将被评估窗口的进展 baseline 绑定到 obligation，并在既有 TS outcome owner 中明确该 trigger 可以接受什么结果。进一步的目标关系修复应通过已有证据关联的 Vision/path 或 successor/blocker 契约完成；不要要求“外部模型再次说好”才能 ACK。回归必须包含真实 refresh：同 observation/同 evidence 被拒绝，有实际新证据的合法处理可结清，同 Turn 重试不会多次结清。

2. [P1] 两个 drift signal 目前不是同一产品语义，不能只当可替换的评分方式

noul_drift_signal 将 behavior_change >= threshold 直接判为非漂移，并且不使用已经采集的 evidence_increment。使用合法响应、阈值 0.6 重放 decoder：

输入事实	Noul drift	Choice drift
行为变化 0.99、服务目标 0.01、新证据 0.01；Choice=off_goal/no_new_evidence	false	true
行为变化 0.01、服务目标 0.01、新证据 0.99；Choice=unknown/new_evidence	true	false

第一行表明做了实质性但无关的改动也可逃过默认规则；第二行表明有新证据、目标关系尚待核实时，仍可被直接标成漂移。这里证明的是规则的确定性差异，不是 live 模型错误率。

建议先分清三件事：目标关系、相对上一窗口的有效增量、有价值的目标内探索/前置工作。behavior_change 只能作为增量证据之一，不能单独证明方向正确；“有新证据”也不能自动洗白无关工作，必须看其与当前问题的关系。把 off-goal、on-goal-but-stalled、useful-exploration、legitimate-wait、unknown 分开，不要压成一个可被 provider 名称切换含义的 bool。Jev 的 Choice/Noul 和原始概率可留在 provider 诊断中；核心消费统一的有限语义事实。

3. [P2] receipt 彼此版本相同，不等于仍适用于当前基准

trigger 只比较被计数 receipts 之间的 contract_revision；context loader 没有读当前评估基准版本。

复用 PR 的真实 capture/refresh/drain fixture（注入模型答案）后，我先生成两份 completed receipts，再修改 observer basis 的 objective/acceptance。当前 basis hash 已改变，但读回的旧 receipts 仍是 completed，仍生成完全相同的 required obligation。现有版本测试只覆盖“一新一旧 receipt 不能凑成连续两轮”，未覆盖“基准已变、尚无新 receipt”的窗口。

assist 需要把 operator study basis 与已生效 Goal/acceptance 基准的绑定讲清楚，并在读取/使用时比对当前 revision。旧 receipt 可保留作历史，不应继续生成当前义务。回归应覆盖已完成结果之后的基准变更、推理过程中变更、晚到结果和 scope 变更；不能只靠 observer 在发送前后检查。

4. [P2] 共用“有效工作轮次”，不要让 bookkeeping 行截断判断

新 scanner 对首个无 receipt 的 run 直接 break，没有使用原 replan 的 neutral-classification 规则。给上述真实 runs 前加入不带 turn identity 的 quota_slot_voided bookkeeping 行后，外部 obligation 从 required 变为 None；同一类型在既有 replan policy 中属于 neutral。这是对生产 reducer 的合成边界测试，没有声称完整执行了 quota-void journey。

应复用有类型的工作轮次投影，区分配额记账、同 Turn 重试、真实未评估的工作轮次。真正 unknown 可以阻止形成新的连续证据，但不能被描述成进展正常，更不能默默充当处理既有义务的 ACK。补上 interleaved spend/void、异步 receipt 未到达和两个 Agent 交错历史的测试。

建议伴随的 TS 重构

不建议本 PR 顺手迁移整个 planner。可以做一个完整而有限的切片：先用 characterization 固定现有 run-history repeat、monitor、periodic-review 的语义和优先级，再把这些历史触发规则与外部 receipt 触发统一到 work_items 下的 TS 纯决策边界。Python 留下历史/receipt 读取和调用适配；原有 Goal-frontier 领域规则继续由原 owner 产生事实，不复制一遍。

触发事实用 discriminated union 表达各自必需字段：重复进展需要 baseline，monitor 需要 target/due facts，外部评估需要基准版本、工作身份和证据窗口。共享 scope、logical-turn 去重、身份与 ACK 匹配，保留各领域不同的阈值、等待条件和合法解除结果。status 与 writeback 调用同一边界；不要为每个 trigger 添加一次独立 RPC，也不要保留第二份 Python 规则。

这样重构的收益是消除重复的身份/时效/历史解释，后续 File、SQLite、PostgreSQL 提供相同事实即可使用相同策略；它本身不等于完成 provider 默认切换。

对主干的风险

default-off 和模型调用放在外部 consumer 是好的隔离措施。但 assist 的 required=true 已经是影响下一步执行的义务，不能仅用 advisory_evidence_only 暗示它没有控制效果。要区分“模型无直接写权限”与“用户启用的确定性策略可以据此要求复盘”。核心 action 文案也应保持 Goal 通用性，避免把“修改文件/改变运行行为”写成所有研究、写作或运营 Goal 的进展标准。

本次执行：

• exact-head tests/capabilities/test_progress_review.py、tests/control_plane/test_external_progress_review.py、packages/loopx-jev/tests：97 passed，包含已有 off/shadow/assist、真实 refresh/status/ACK 测试。
• 额外隔离反例：上述原样 ACK、信号差异、完成结果后的 basis 变更、neutral bookkeeping 都得到所述结果。
• git diff --check 在 7 个 fixtures/sentinel/real/*/COMMIT.txt 报新增 EOF 空行，未把它报告为通过。
• 未运行新的付费 live inference、完整 frontend packaged journey、全部 CI 或 provider 切换验收；这些不是本次设计反馈已经通过的证据。

后续对照需要包含有正常上下文的现有 Agent/replan 流程，而不仅是精确重复 fuse。衡量首次正确且可行动的触发时间、误触发率、合法探索损失、复盘消耗与触发后是否恢复有效推进。连续两次同模型判断不等于两份独立证据，不能直接换算成更高置信度。当前 recorded matrix 有实验价值，但还不能证明任务完成收益。

我的整体评价

建议沿“现有 replan 的语义证据增强 + 有界 TS 触发规则收敛”继续。先修复上面的闭环与有效性问题，再讨论 assist 采用；保留 shadow 作为校准入口。代码上可以按同一主题分两个可独立复核的阶段：① 原规则 characterization/迁移，证明 off-path parity；② 新语义事实与有效性/解除契约接入，证明正反闭环。若仍保持一个 PR，也请用这两个逻辑阶段组织 commits 和验证。

不需要为了这个方向重写整个 Goal-frontier、另造 planner 或扩大模型权限。最关键的完成标准是：触发有当前有效证据，处置对应触发原因，确认能证明问题已处理或有证据地保留原计划。

English verdict: REQUEST_CHANGES - design feedback at 6be1537, not a merge-readiness certification. Reuse the existing replan lifecycle and converge the bounded history-trigger policy in TypeScript. Fix the reproduced identical-observation ACK bypass, distinguish alignment from behavior/evidence novelty, bind completed receipts to the current approved basis, and preserve neutral bookkeeping semantics. The 97 focused tests pass, but the additional counterexamples expose gaps outside that coverage; no new live-model quality or full integration approval is claimed.

[songoow]

Revised after the external review of head 6be153723. All four determinism findings are fixed in code rather than by tuning the model, and the matrix was re-recorded with a v2 question set that removes goal identity from the model input:

• Label leak: the harness no longer passes a case-derived goal_id; the model state is a whitelist of operator basis fields, and a test pins byte-identical requests across identities.
• Drift rule semantics: noul now gates on serves_acceptance ∧ evidence_increment, both asked about the change between checkpoints; behavior_change is recorded only. The core recomputes the booleans from the typed judgments and rejects a receipt that disagrees.
• Contract currency: assist requires --progress-review-contract-revision pinned to the observer basis printed by drift init; other revisions are stale, unpinned assist reports contract_revision_unpinned.
• Identity and lifecycle: turn-id joins require Agent/Todo agreement, ambiguous fallbacks are unattributed, and up to two newest pending receipts are skipped so a streak neither grows nor dissolves while evaluation runs.

Re-recorded differential (two independent live runs agree on all 16 sequences): typed fuse 0/9, noul 9/9 flagged at the gold round and 9/9 reaching the obligation, 0/7 false flags on real commits, 0 premature; choice 5/9. Post-implementation churn now scores serves_acceptance 0.06–0.13 (was 0.62–0.90). Caveat kept in the docs: the v2 wording was revised after seeing v1 misses on the constructed sequences, so those are no longer held-out for the wording; the real commits were untouched.

CI: fixed the mypy Callable type, the control-plane→capabilities import edge (policy now lives in control_plane/work_items/progress_review_policy.py), the ratchet oversized function, and repackaged the dashboard assets; rebased onto 0d464d2ac for the upstream goal-channel and SQLite Node fixes.

[huangruiteng]

Exact-head review — e7afab6daedba91aee15a7f312510e50ac12245d

当前结论是 REQUEST_CHANGES。最关键的 P1 不是模型质量问题，而是 control-plane 的 discharge 语义仍有缺口：external_progress_review_drift 没有把已评估的 progress baseline 交给现有 replan 语义，因此完全相同的 observation 会被误判成新进展并解除义务。另有 7 个 fixture 文件未通过 git diff --check。

动机

这个 PR 解决的是一个真实且重要的问题：agent 自报 advanced 并不能证明目标真的前进，连续的改名、格式整理或重复尝试可能掩盖 drift。PR 的目标是引入一个默认关闭、显式启用的 progress-review sentinel：外部 observer 只产生带 contract revision 的 typed receipts；shadow 只观测；assist 达到阈值后，把证据接入现有 autonomous_replan_obligation，但不获得 Todo、lease、quota 或执行权限。

这个边界总体是对的。它复用了现有 typed progress/replan vocabulary，并通过 off/shadow/assist、pinned contract revision 和 advisory receipt 把可用性、激活与权限分开；这比让 provider 直接写 Todo 或另起一套 scheduler/replan 状态机更符合当前架构。不过“触发 obligation”只是闭环的一半，另一半是 obligation 只能被真实的新语义证据解除。本 head 尚未满足这一点，所以完整交付目标还没有成立。

改动思路

主要链路是：loopx-jev 从 Goal basis、代码变化和既有 evidence 构造匿名请求并记录 typed review receipt；core 的 external_progress_review_context 只选择当前 pinned contract 的 receipts；external_progress_review_trigger 把达到阈值的 drift 变成 external_progress_review_drift；autonomous_replan_obligation_from_runs 继续生成既有 obligation；最终 refresh_state_run 通过 semantic_delta_from_writeback 和 TypeScript replan semantics 判断本轮 writeback 能否 ACK。

配置与产品入口也形成了相对完整的路径：capability catalog、configure-goal、chat Goal configuration API、dashboard localization/packaged assets、双语 README/能力文档，以及独立的 packages/loopx-jev provider。默认 off；仅安装 package、存在 credential、worker 可用或已有 receipt 都不会激活 assist；shadow 不产生机器义务；assist 还要求 contract revision 被 pin。这些 counterfactual 在测试里有直接覆盖。

问题出在 core 复用边界的最后一步。现有 typed_progress_repeat trigger 会携带 progress_baseline，所以相同 surface/hypothesis/evidence 会被判为 repeat；新的 external trigger 只携带 review metadata 和 frontier_identity，没有 baseline。下游没有足够状态区分“触发 obligation 的旧 observation”和“本轮真正的新 observation”。

具体改动

• loopx/capabilities/progress_review/**、catalog、Goal 配置 API 与 dashboard 增加 policy/config/readback，公开 off、shadow、assist 与 contract-revision pinning。
• loopx/control_plane/work_items/external_progress_review.py 与 progress_review_policy.py 负责读取/筛选 typed receipts 并形成 goal-neutral trigger；status.py、state_refresh.py 和 obligation projection 把它接入现有控制面。
• packages/loopx-jev/** 实现可选 observer 的配置、协议、runner、HTTP worker、drift capture、sentinel matrix 与 receipt store，并附带 constructed/real fixtures 和 recorded responses。
• README、双语 capability/RFC/package 文档说明激活、隐私、authority 和 rollout；tests 与 public smoke 覆盖默认关闭、shadow、assist、contract mismatch、matrix 结果和一个正向 ACK/re-arm 流程。

关键代码讲解

1. external_progress_review_trigger：把当前 contract 的 drift receipts 压缩为一个 typed trigger。它正确保留 signal、threshold、evidence ids 和 contract revision，但没有保留触发窗口对应的 progress observation baseline。
2. autonomous_replan_obligation_from_runs：统一拼装 typed repeat、dead monitor、blocked successor 与 external review 的 obligation。当前代码只在 typed_progress_repeat 分支写入顶层 progress_baseline；随后 external review 分支仅覆盖 frontier 和 review metadata。
3. semantic_delta_from_writeback / semantic_progress_delta：这是实际 ACK gate。前者从 obligation/trigger 读取 baseline；读取不到时传 None，后者会把一个完整 observation 当成首次出现，于是相同 surface 与 hypothesis 被分类为 new_surface / new_hypothesis。
4. test_same_sequence_off_sees_nothing_and_assist_raises_the_obligation：很好地证明了 off/shadow/assist 与正向 re-arm，但 ACK 使用的是全新的 blocked observation，因此没有覆盖“把刚刚触发 drift 的 observation 原样提交”这个负向路径。

对主干的风险

P1 — 相同 observation 可以错误解除 external drift obligation

触发状态：assist 已因两个 current-contract drift receipts 生成 external_progress_review_drift，agent 随后提交与已评估 run 完全相同的 surface_id、hypothesis_id 和 evidence_ids。

代码路径：loopx/control_plane/work_items/autonomous_replan_obligation.py 的 baseline 复制逻辑只匹配 typed_progress_repeat；external review 分支没有写 baseline。loopx/control_plane/work_items/progress_observation.py 随后以 baseline=None 调用 semantic_progress_delta。

我在 exact head 上用生产函数复现到：trigger 是 external_progress_review_drift，obligation 不含 baseline，完全相同的 observation 返回 accepted: true，outcomes 为 new_surface、new_hypothesis。这会让 agent 在没有产生新证据、blocker、successor 或 terminal boundary 的情况下 ACK 机器义务，状态表面显示“已恢复”，实际 drift 没有停止。

最小修复：把 external review 所评估窗口的 typed progress baseline 绑定到 trigger/obligation，或者在既有 TypeScript semantic owner 中定义等价的 trigger-specific discharge rule；不要把 ACK authority 下放到 loopx-jev。需要补一个真实 refresh_state_run 回归：相同 observation 必须拒绝；真正的新 evidence、blocker、successor 或 terminal outcome 分别允许。建议的最小重跑集合是相关 control-plane test、packages/loopx-jev/tests/test_closed_loop.py、typecheck 与 public sentinel smoke。

P2 — exact diff 仍未通过 whitespace gate

git diff --check origin/main...HEAD（以及 latest-main staged integration）报告 7 个 packages/loopx-jev/tests/fixtures/sentinel/real/*/COMMIT.txt 文件新增 EOF 空行：docs_vision_schema_compaction、fix_closeout_preflight_latency、fix_lark_part_sequence_settlement、fix_manager_refused_read_argument、fix_settled_turn_safe_bypass、test_closeout_preflight_budget、test_registry_smoke_external_evidence。请清理后重跑 git diff --check。

语义与 CI 对齐

本地 exact-head 验证中，104 个 focused tests、npm run typecheck:control-plane 与 16-case public smoke 都通过；exact head 也能干净合入最新 origin/main。但这些绿灯没有覆盖上述 negative transition，而当前远端 CI 仍有运行中的 Python shards，早先 superseded run 还出现过 merge-gate failure。packet 对本 Goal 明确是 wait_for_ci=false，所以本评审没有等待 CI；结论由已复现的语义反例决定。

此外，整份变更为 170 files、+17,311/-10，主体体量来自 optional provider、real/constructed fixtures 与 34 份 recorded responses。core mechanism 本身是合理的；修掉 blocker 后仍建议做一次有证据的 durability pass：保留能保护 shipped contract 的最小矩阵，把不承担长期 regression contract 的研究记录压缩或归档。这一点目前作为维护性建议，不替代 P1/P2。

我的整体评价

方向和 authority 边界是扎实的：默认关闭、shadow/assist 分层、contract pinning、anonymous request、typed receipt、复用既有 obligation/TypeScript semantics，均比新建平行权威更稳妥；配置、UI、CLI、状态 readback 和双语文档也基本形成了完整用户路径。未来向前的 bounded refactor 应继续放在现有 replan owner：让所有 trigger producer 提供统一的 frontier/baseline contract，避免每个 provider 自己解释“新进展”。

但当前 exact head 的核心闭环仍可被重复 observation 绕过，这不是非阻塞建议；再加上确定的 diff-check 失败，我不能批准。请修复 typed baseline/discharge invariant、加入真实负向回归并清理 7 个 fixture EOF；保持同一 head 重新跑 focused suite、typecheck、public smoke、git diff --check 和 latest-main integration 后再复审。没有执行 merge。

English verdict: REQUEST_CHANGES - at e7afab6, external_progress_review_drift lacks the baseline needed to reject an identical observation, so unchanged work can falsely ACK the enforced replan obligation; seven fixture files also fail git diff --check, while 104 focused tests, typecheck, the 16-case smoke, and latest-main integration otherwise passed.