Preserve Turn loading isolation and qualify scoped interruption recovery

docs/architecture/rfcs/monorepo-distribution-split-v0.md

@@ -2699,6 +2699,98 @@ context/model/split-install/PostgreSQL-host/full-cost/adoption gaps remain open.
 The original task and independent review/readiness remain open; no main merge
 or installed-host upgrade is included.
 
+### 2026-11-08 — Report construction loading at Turn composition
+
+The existing `periodic-report` Python provider and CLI Turn adapter now resolve
+report source, preset, stage projection and request exports on demand. At runtime
+revision `7858d96e83ef98174cc43870e283f8f0e6f40c0b`, disabled subscriptions do
+not import those construction providers. The capability package already had lazy
+exports; this change fixes its concrete composition callers. Subscription
+validation remains with the same owner, including its cadence-to-profile schedule
+validator. Goal overrides, live machine defaults, configuration warnings, original
+export identities and patch points, hook profile digest/policy version, source
+selection, receipt replay and authority are retained. It adds no capability,
+configuration, UI, persisted vocabulary or decision owner.
+
+Against baseline `c263f7c78353fcaa914f23475d4d99d07eab7c01`, eleven identical
+synthetic configuration cases preserve hook output, trigger decisions and complete
+configuration error messages. The cold-process refusal oracle fails before the
+change when disabled composition imports the report source. Eight focused cases
+now cover absence, disabled inheritance, overrides, invalid disabled configuration,
+effect-free reads and export/patch compatibility. An initial stronger prohibition
+also rejected the canonical profile schedule validator; that failed probe is
+retained, and this slice preserves the validator rather than bypassing it.
+
+Source and separately built wheel File/default-SQLite console journeys pass the
+original complete 101-Todo lifecycle: unchanged fixture/artifact, four receipts,
+one debit, source/lease/refusal/restart/replay readback and two owned native exits
+per journey. Installed package ownership was verified before comparing 1,920
+present tracked entries; seventeen source-only or documentation/media omissions
+remain explicit. Wheel test-dependency and frontend-build preparation failures
+remain recorded. Isolated PostgreSQL 16 store/service suites pass 391/11 cases
+with zero skips; the initial service run's missing environment declaration remains
+a skipped run. This does not qualify PostgreSQL host/wheel journeys. Report suites
+pass 100 cases. The broader digest inventory check has one identical baseline/head
+failure on an unchanged Todo selector; that check remains failed.
+
+A separate report-composition probe used CPython 3.13.5 on macOS arm64, normal
+bytecode in both arms, an uncounted prewarm and eight alternating pairs per
+disabled/enabled case. Disabled median wall/CPU observations were 0.255/0.111 s
+before and 0.246/0.107 s after; enabled observations were 0.331/0.117 s and
+0.321/0.113 s. Loaded LoopX modules fell from 196 to 179/182. These small,
+concurrent-test observations do not establish material speedup, sustained tail
+latency or complete CLI/Node/model/attention cost benefit. The earlier frozen
+20-cost, attribution and normal-bytecode campaigns were not repeated, enlarged or
+relabeled. The independently reversible delta is off-state construction isolation.
+Exact-head review, CI/readiness and explicit protected merge authority still govern
+adoption. Preserve I3 for at least one minor release, M1/#8011, full R0/V10/R1,
+D2/D5/D6/D3, context/model/split-install, full-cost/adoption gaps and all historical
+failures. The existing runtime-composition work continues through these boundaries.
+
+### 2026-11-08 — Synthetic Host UTF-8 recovery qualification
+
+The Windows CPython 3.11 CI arm of #8360 fails six existing original-task
+revalidation cases. Its synthetic Python Host inherits a locale decoder for a
+pipe, while the existing generic Host transport sends UTF-8 JSON. A disposable
+File-provider counterprobe changes only that Host decoder: CP1252 produces
+`host_failure` without `task_failure`; UTF-8 reaches the independent artifact
+validator and produces `validation_failed` at `task_postcondition`, with
+`revalidate_cached_result`. The source macOS arm independently passes all 12
+existing revalidation cases. These observations locate a fixture mismatch;
+they do not justify widening production failure classification.
+
+Fixture commit `a4ce613585fdb329ba3600f0ede768864214ed84`, over main
+`4e949f405528fd8087f0623526c62f890c1ce982`, explicitly configures its stdin as
+UTF-8 before reading the request. One new File/SQLite regression first sets
+only the generated Host's decoder to CP1252, sends a Unicode brief and requires
+the rejected artifact to reach task validation. After repairing the artifact,
+revalidation must accept it with the original turn key and exactly one Host
+invocation. The same regression fails with the immutable baseline Host on the
+same CPython 3.13.5/Node 22.22.3 environment, and passes with the correction.
+No transport, production owner, persisted state or protocol vocabulary changes.
+
+On macOS arm64 with Node 22.22.3, CPython 3.11.15 source and a newly installed,
+isolated monolithic wheel each pass the new two provider cases plus all 12
+existing original-task recovery cases. Wheel imports are verified outside the
+checkout before loading the source-owned fixtures; all 1,920 present packaged
+tracked runtime entries match the source bytes. The same 17 source-only
+exclusions remain: canary baseline, host plugin/config, fault-test helper and
+documentation/media. This qualifies the existing File/SQLite
+synthetic Host recovery path, including retained MCP disconnection and lock
+timeout recovery cases, not model adoption, split installation or native
+Windows. The failed original Windows arm remains historical evidence until a
+new exact-head Windows run passes. The independent Frontstage manifest request
+still returns HTTP 504 after its bounded retry.
+
+The original task stays open. Preserve I3 for at least one minor release,
+M1/#8011, full R0/V10/R1, D2/D5/D6/D3, context/model/split-install,
+PostgreSQL host/wheel, whole-path/full marginal-cost/adoption requirements and
+all historical and frozen failures. No cost campaign is repeated or relabeled,
+and this fixture qualification establishes no performance benefit. The report
+loading proposal #8360 retains its independent review and CI gates; the next
+step is exact-head review of this correction and native Windows readback. No
+main merge or installed-host upgrade is included.
+
 ## Appendix B: Decision log
 
 | Date | Decision | Owner / approval | Alternatives | Normative sections changed |
@@ -2822,3 +2914,44 @@ I3/minor-release, M1/#8011, full R0/V10/R1, D2/D5/D6/D3,
 context/model/split-install and historical failure qualifications remain open.
 The next useful boundary remains report subscription/construction loading after
 the combined proposal's review; the original task stays open.
+
+### 2026-11-08 — Integrated loading and recovery qualification
+
+The combined proposal adopts the existing feedback/declaration-codec repair
+from #8346, report construction isolation from #8360 and UTF-8 synthetic Host
+fixture repair from #8365 over target main `fe36f31fd`. Runtime qualification
+uses `251d7eaeff4581d3436340f68c5b6cd8c8ebf1ad`. The existing lexical
+`write_scope_codec.ts` already removes the observed dependency on full lease
+admission; an additional interaction decoder layer is unnecessary. Error and
+vocabulary owners, declared scope spellings and retained receipts stay intact.
+
+On Node 22.22.3, the loading/digest checks and adjacent work-requirement,
+lease and effect-program regressions pass (60 cases), as does the control-plane
+typecheck. Selected-owner load failure still rejects without fallback. Four
+fresh source/wheel × File/SQLite 101-Todo console correctness episodes preserve
+the prior immutable fixture/artifact fingerprints, scopes, original completion
+time, four receipts, one debit, refusal/restart/replay and two owned native exits
+each. All 1,921 present tracked wheel entries match source bytes; seventeen
+source-only, documentation and media omissions are unchanged.
+
+Real PostgreSQL 16.15 store/service suites pass (391/11, zero skips) on an
+isolated disposable server that was stopped afterward. The first service run
+omitted its dedicated environment variable and skipped one case; that result
+remains a failure of qualification, separate from the corrected run.
+
+A browser-fixture interruption check also depended on the shared five-second
+completion window and checked the request ledger before its asynchronous receipt
+returned. The scenario now holds only the correction's exact Session/Turn stream,
+checks that it is still active and waits for the matching interrupt response.
+The original ten-second interrupted-message assertion remains intact. A delayed
+receipt probe fails the original scenario; removing only the stream hold from the
+candidate reproduces that message timeout, while the complete candidate passes.
+This is synthetic browser qualification, not proof of a native-runtime defect or
+an installed repair. Completion priority and stale-Turn guards remain unchanged;
+the combined head still requires packaged validation, CI and independent review.
+
+These episodes qualify integration correctness. Whole-path and full marginal
+cost and adoption remain unqualified. Preserve I3 for at least one minor
+release, M1/#8011, full R0/V10/R1, D2/D5/D6/D3, context/model/split-install,
+PostgreSQL host/wheel and every frozen failure. The original task remains open;
+this checkpoint does not authorize merge or declare the minimal kernel complete.

docs/architecture/rfcs/monorepo-distribution-split-v0.zh-CN.md

@@ -2096,6 +2096,79 @@ hard-cut 证据、I3／minor-release、M1/#8011、完整 R0/V10/R1、D2/D5/D6/D3
 context／model／split-install／PostgreSQL-host／full-cost／adoption 缺口保持开放。
 原 Task、独立评审／readiness 继续开放，本轮不合入 main、不升级 installed host。
 
+### 2026-11-08 — Turn 装配中的报告构造加载
+
+既有 `periodic-report` Python provider 与 CLI Turn 适配层改为按需解析报告源、
+预设、阶段投影和请求导出。在 runtime 版本
+`7858d96e83ef98174cc43870e283f8f0e6f40c0b`，禁用订阅不会导入这些构造 provider。
+能力包已有 lazy exports，本次修复的是实际装配调用方。订阅校验仍由原 owner
+负责，包括 cadence 使用 profile 中的日程校验器。Goal override、live machine
+defaults、配置 warning、原导出身份与 patch 点、hook profile digest/policy version、
+source 选择、receipt replay 和权限保持；不新增 capability、配置、UI、持久化词汇
+或决策 owner。
+
+与基线 `c263f7c78353fcaa914f23475d4d99d07eab7c01` 对照，11 个相同合成配置场景
+保持 hook 输出、trigger 决策及完整配置错误消息一致。冷进程拒绝 oracle 在修复前
+因禁用装配导入 report source 而失败；8 个聚焦场景现覆盖缺省、禁用继承、override、
+非法禁用配置、无副作用读取和导出／patch 兼容。最初更强的禁止规则也拒绝了原
+profile 日程校验器；该失败探针保留，本阶段保持校验器，不绕过它。
+
+Source 与独立构建 wheel 的 File/default-SQLite console 路径通过原始完整
+101-Todo 生命周期：fixture/artifact 不变、4 个 receipts、一次 debit、source/lease/
+拒绝／重启／重放读回，以及每条路径 2 个自有 native 退出。先验证安装 package
+归属，再比对 1,920 个实际存在的 tracked entries；17 个 source-only 或文档／媒体
+省略明确保留。Wheel 测试依赖和 frontend build 准备失败仍记录为失败。
+隔离 PostgreSQL 16 store/service 套件通过 391/11 场景，零 skip；service 首轮
+因缺少环境声明而跳过的记录仍保留。这不资格 PostgreSQL host/wheel 路径。
+报告相关套件通过 100 场景；较广的 digest inventory 检查在未改动 Todo selector
+上出现一条完全相同的 baseline/head 失败，该检查仍为失败。
+
+独立报告装配探针在 macOS arm64/CPython 3.13.5 下，两侧使用正常 bytecode、一次
+不计样本的预热，对禁用／启用场景各做 8 对交替观察。禁用 wall/CPU 中位数为
+修复前 0.255/0.111 s、修复后 0.246/0.107 s；启用为 0.331/0.117 s 和
+0.321/0.113 s。加载的 LoopX modules 从 196 降到 179/182。小样本与并行测试下
+的这些观察不证明明显提速、持续尾延迟或完整 CLI/Node/model/attention 成本收益。
+早先冻结的 20-cost、归因及正常 bytecode 实验未重跑、扩充或重标签。
+本阶段可独立回滚的增量是关闭态构造隔离。采纳仍需同 head review、CI/readiness
+和明确的受保护合并授权。I3 至少保留一个 minor release，M1/#8011、完整
+R0/V10/R1、D2/D5/D6/D3、context/model/split-install、full-cost/adoption 缺口
+及所有历史失败均保持开放；既有 runtime composition 工作继续跨越这些边界。
+
+### 2026-11-08 — 合成 Host 的 UTF-8 恢复资格
+
+#8360 的 Windows CPython 3.11 CI arm 在六个既有原任务重验证用例中失败。
+合成 Python Host 对 pipe 继承 locale decoder，而原 generic Host transport
+发送 UTF-8 JSON。Disposable File-provider counterprobe 只改变该 Host decoder：
+CP1252 产生 `host_failure`，没有 `task_failure`；UTF-8 到达独立产物 validator，
+在 `task_postcondition` 产生 `validation_failed`，并提供
+`revalidate_cached_result`。源码 macOS arm 独立通过全部12个既有重验证用例。
+这些观察定位了 fixture mismatch，不支持扩大生产 failure classification。
+
+Fixture commit `a4ce613585fdb329ba3600f0ede768864214ed84` 基于 main
+`4e949f405528fd8087f0623526c62f890c1ce982`，在读取请求前显式将 stdin 配为
+UTF-8。新增 File／SQLite regression 先只把生成 Host 的 decoder 设为 CP1252，
+发送 Unicode brief，要求被拒绝的产物到达 task validation。修复产物后，重验证
+必须沿用原 turn key 并接受结果，Host invocation 始终为一次。同一 regression
+在相同 CPython3.13.5／Node22.22.3 环境中对不可变 baseline Host 失败，修正后
+通过。不改变 transport、生产 owner、持久状态或协议词汇。
+
+macOS arm64／Node22.22.3 下，CPython3.11.15 source 与新安装的隔离单体 wheel
+各通过新增两个 provider 用例及全部12个既有原任务恢复用例。先在 checkout 外
+验证 wheel import 归属，再加载 source-owned fixture；1920个实际 packaged
+tracked runtime entry 均与源码字节相同。原17个 source-only exclusion 保留：
+canary baseline、host plugin／config、fault-test helper 和 documentation／media。
+这验收既有 File／SQLite 合成 Host 恢复路径，包括保留的 MCP
+断连及 lock-timeout 恢复用例，不验收模型采纳、split installation 或原生 Windows。
+原失败 Windows arm 保留为历史证据，直到新 exact-head Windows run 通过。
+独立 Frontstage manifest 请求在有界重试后仍为 HTTP504。
+
+原任务保持开放。保留 I3 至少一个 minor release、M1/#8011、完整 R0/V10/R1、
+D2/D5/D6/D3、context/model/split-install、PostgreSQL host/wheel、whole-path／
+full marginal-cost／adoption 验收及全部历史／冻结失败。不重跑或重标签成本
+campaign，本 fixture 资格不证明性能收益。报告按需加载提案 #8360 保留独立
+review／CI gate；下一步为本修正的 exact-head 评审与原生 Windows 读回。不包含
+合入 main 或 installed-host 升级。
+
 ## 附录 B：决策日志
 
 | 日期 | 决策 | 负责人 / 批准 | 备选 | 变更的规范章节 |
@@ -2191,3 +2264,35 @@ Whole-path 成本收益、installed adoption、PostgreSQL host／wheel，以及
 I3／minor-release、M1/#8011、完整 R0/V10/R1、D2/D5/D6/D3、
 context／model／split-install 与历史失败资格继续开放。组合提案评审之后，下一段
 有用边界仍是报告订阅／构造加载；原 Task 保持开放。
+
+### 2026-11-08 — 组合加载与恢复资格验证
+
+组合提案在目标 main `fe36f31fd` 上采用 #8346 已有的 feedback/声明 codec
+修复、#8360 的报告构造隔离和 #8365 的 UTF-8 合成 Host fixture 修复。
+运行时代码资格基于 `251d7eaeff4581d3436340f68c5b6cd8c8ebf1ad`。
+已有的词法 `write_scope_codec.ts` 已解除观察到的完整 lease 准入依赖，无需
+增加 interaction decoder 层。错误与词汇 owner、声明 scope 拼写和保留回执均保持。
+
+Node 22.22.3 上的加载/digest 与相邻 work-requirement、lease、effect-program
+回归全部通过（60 项），控制面类型检查通过。选中 owner 的加载失败仍拒绝执行，
+不会回退。四次新的源码/wheel × File/SQLite 101-Todo console 正确性路径均保留
+前次不可变 fixture/artifact 指纹、scope、原完成时间、四份回执、一次扣额、
+拒绝/重启/重放，以及每条路径两个自有 native 进程退出。wheel 中现存的 1,921 个
+tracked 条目逐字节匹配源码，十七项源码专用、文档和媒体遗漏保持不变。
+
+真实 PostgreSQL 16.15 store/service 在隔离临时服务上全部通过（391/11，零
+跳过），自有服务随后停止。首次 service 运行漏传专用环境变量，跳过一项；
+该次资格失败单独保留，与修正后的运行区分。
+
+浏览器 fixture 的中断检查还依赖共享的五秒完成窗口，并在异步回执返回前检查请求
+记录。该场景现在只保持纠偏所对应的准确 Session/Turn 事件流，确认它仍在执行，
+并等待匹配的中断响应。原十秒中断消息断言完整保留。延迟回执探针使原场景失败；
+仅移除候选中的流保持、保留响应等待，会复现该消息超时，而完整候选通过。
+这是合成浏览器资格验证，不证明 native runtime 存在缺陷或已安装修复。
+完成优先级与 stale-Turn 保护保持不变；组合 head 仍需 packaged 验证、CI 和
+独立评审。
+
+这些路径验证组合正确性，whole-path 与完整边际成本、adoption 仍未取得资格。
+保留 I3 至少一个 minor release、M1/#8011、完整 R0/V10/R1、D2/D5/D6/D3、
+context/model/split-install、PostgreSQL host/wheel 以及每项冻结失败。
+原任务保持开放；本检查点不授权合入，也不宣告最小化内核完成。