fix(cli): reuse canonical bootstrap for source invocations

[huangruiteng]

What this delivers / 本次交付

Source callers (python -m loopx.cli) now reuse the existing console bootstrap instead of importing every command owner before dispatch. The imported loopx.cli.main / build_parser API remains the full-parser compatibility surface.

源 CLI 复用现有 console bootstrap，消除先加载全部命令 owner 的入口漂移；库方式调用保留完整 parser。这是 Python 宿主适配器修复，复用现有 TS/command owners，不增加 Python 策略源、命令目录或 provider。

The bound scheduler route also preserves the existing outer-controller write guard: an outer-controlled invocation cannot process-replace the dispatcher before that guard runs. Missing Node on a normally bound TS route still fails closed; manual/unbound compatibility remains.

已绑定 scheduler 路线保留外层控制器写保护，不能在 guard 执行前 process-replace；正常 TS 路线缺少 Node 时仍关闭失败，手工/无绑定调用继续兼容。

Changed surfaces / 变更范围

• loopx/cli.py: route only module execution through loopx.entrypoint.main; retain imported full-parser APIs and canonical fallback diagnostics.
• loopx/entrypoint.py: keep outer-controlled calls in the dispatcher that already owns Goal matching and write denial.
• tests/control_plane/test_source_cli_entrypoint.py: 32 fresh-process cases covering owner loading, parser fallback, TS routing, no-Node failure, guard ordering, usage disclosure/opt-out, real File/SQLite reads, independent-Goal scope and actual scheduler-write recovery.
• docs/reference/source-cli-entrypoint.md: bilingual ownership, compatibility, user journey, qualification and honest timing boundaries.

Validation and limits / 验收与局限

• At exact head 440cefcfca376c03190384d8546658e912b5fc8e, based on main 6643f367064b9921c864db75a042979fddc4b8c3 (includes separately merged perf(contract): overlap bounded full-tree file reads #5239), the full source/CLI/usage/completion/scheduler/native-controller/public-reader matrix passed: 168 passed, 1 existing warning, 165.59 s, with no skips. This includes all 32 source-entry cases. The two supplementary source/console scope-and-recovery cases also passed independently before the final matrix.
• Ruff, compile checks and configured mypy (19 source files) passed; whole diff whitespace checks passed.
• Independent counterfactual consumers rejected the old eager source entry, deliberately reintroduced unrelated-owner loading and removal of the outer-controller route guard. Mutations were restored before committing.
• Real File/SQLite reads retained exact canonical provider revisions with Markdown display removed, with no authority mutation or recovery from that display.
• Real receipt-bound scheduler tests reject the protected Goal even if the actor changes, allow both existing and newly registered independent Goals, and commit/read back the original Goal's scheduler state after its controller releases the guard. Ordinary health-failure reads are distinguished from guard rejection; this does not claim that an unrelated Goal's business task or health is repaired.
• The normal packaged-frontend build and bundle verification passed, followed by a wheel build and an isolated wheel installation outside the source tree. Installed source/console version, help, canonical parser rejection, pure JSON and protected-write/no-state checks passed (12 invocations). The wheel contains the packaged chat assets and TS follow-up owner; installed bootstrap files exactly match this head. The initial wheel attempt correctly refused a missing prebuilt chat bundle; the normal build supplied it without bypassing the packaging gate. No frontend source changed and the live installed runtime was not replaced.
• One earlier pre-rebase large run had 151 passed / 1 unchanged usage timing assertion failed (foreground 2.817 s vs baseline 1.650 s + 1 s). The assertion was not changed. Four independent alternating base/candidate reruns of that case passed (3.45, 3.80, 3.08, 2.95 s pytest duration); the latest full matrix also passed. Shared active-host load is a limitation, not proven failure attribution.

中文：完整回归、独立入口/权限反例和真实 File/SQLite 读回均已执行。保留此前耗时断言失败及复验结果，没有放宽断言、删除验收或把 host load 推断当作根因证明。

This does not close the whole-batch deadline task. Pre-rebase ABBA runs of the same unchanged 28-case scheduler matrix all passed but took base 129.28/123.24 s and candidate 120.63/191.03 s. The adverse candidate sample does not establish a stable whole-batch improvement; the original 20-second budget remains unmet and unchanged. The separately merged #5239 is not retroactively included in those measurements.

不宣称整批期限已达成。 28 项原断言全部保留，但候选有变慢样本，完整负载仍远超 20 秒。继续保留任务与验收缺口，不把启动改进算作 terminal 完成，也不扩大 validator deadline。

An alternating fresh-process startup probe on the shared active host measured --version medians 0.759 → 0.040 s and quota --help 0.805 → 0.072 s. These are tiny/help startup measurements only, not business operation latency, isolated cold-machine cost or p95.

Product path and compatibility / 产品路径与兼容性

Affected: source CLI and managed/canary callers invoking that module. Source now intentionally shares console help and existing usage policy: first eligible use may disclose on stderr; JSON stdout stays pure; help/version do not observe; existing opt-outs and machine settings are preserved. This is not byte parity with the old source entry's absence of disclosure.

受影响入口为源 CLI 及调用它的 managed/canary；首条使用告知和机器设置统一到现有 owner。前端/Lark 不新增控件或配置/schema，继续消费相同投影与回执；既有 usage-settings HTTP roundtrip 已验收。本次没有前端源码或布局变化；为验证完整 wheel，已按既有流程生成并校验打包资源，不冒充新前端/Lark transport E2E 验收。

The installation above is an isolated artifact check only, not live deployment or provider promotion. The TS migration RFC's single-owner boundary remains; core merge stays with maintainers.

没有宣称已安装、provider 晋级或 TS 全量切换；LoopX core 合并仍交维护者。

[huangruiteng]

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

Exact head: 440cefc; base: 6643f36.

动机

没有阻塞性发现；本结论只批准可独立验收的入口修复，不批准整批期限任务完成。真实 managed/canary caller 会执行源模块，旧源入口先导入所有命令，偏离 console 的轻量启动路径；console 的已绑定 scheduler follow-up 还可能在外层控制器 guard 前替换进程。本次把入口收敛回现有 owner，减少无关加载并恢复该权限边界。原多命令验收的 20 秒预算仍未达成，后续成本优化继续由原任务承担。

改动思路

源模块执行先进入既有 entrypoint，普通常用命令交给 selected dispatcher，拒绝参数及其他命令回到完整 parser，库方式导入的完整 API 保留。已绑定 follow-up 正常复用现有 TS executable；outer-controller 标记存在时不提前替换进程，让既有 guard 解析实际 Goal 并判断写权限。标记只选择宿主路线，不新增权限事实，也不复制 Goal 匹配规则。缺少 Node 的正常绑定路线仍关闭失败。此边界符合 TS 重构 RFC 的单一 owner 原则：Python 只做宿主适配，TS 与原控制面继续负责调度、状态、回执和结算。

具体改动

全量 diff 为四个文件，508 行增加、4 行删除；生产部分只有14行增加、4行删除，其余为337行回归和157行双语说明。测试不是只比较两个新 helper：新解释器实际执行源模块，与 console 比较加载、完整诊断及输出；真实 File/SQLite 读回在移除 Markdown 展示后仍保留原 provider revision，且不改权威状态。追加的两个范围测试覆盖读取、现有及后注册的独立 Goal、改变 actor 不能绕过保护，以及 controller 解除保护后的真实 ACK 提交与单独读回。fixture 无业务 Todo 时允许普通 health-failure 读取，不把它包装成业务恢复。

关键代码讲解

1. loopx/cli.py:7 的 module-execution 分支在 command imports 前调用现有 bootstrap；按名字再次导入完整 parser 时不会重入此分支，避免递归并保留库 API。
2. loopx/entrypoint.py:20 的 _native_scheduler_followup_argv 在外层控制器标记为精确值1时停止提前 native dispatch。它不做 Goal 权限判断，实际规则仍在既有 guard，正常绑定及缺 Node 分支保留。
3. loopx/cli_runtime.py:344 的既有 _dispatch_selected 先执行 guard，再解析 registry、进入 handler；新源入口复用它，拒绝发生在 provider effect 前。
4. loopx/cli_runtime.py:360 的既有 _run_full_cli 按名字导入完整 API，selected parser 拒绝的参数恢复原诊断，未新增第二套 grammar。

对主干的风险

最强反例是“禁止绕过但误拦所有 Goal”或“返回 blocker 却无法恢复”。真实 module/console 反例均通过：受保护 Goal 即使 actor 改变仍拒绝且无 scheduler state；同一 registry 的独立 Goal 及启用后新 Goal 可实际写回；原 controller 解除保护后原 Goal 也能提交并读回。这证明调度写入恢复，不证明整段金融工作已交付。独立 base/head probe 还能识别旧 eager 加载和提前 exec；exec 截获本身不是实际 TS 写入证明，真实回执/state 测试补足该限制。

源入口现在有意采用既有 console help 和 usage-policy：首条合资格命令可能在 stderr 告知，JSON stdout 不受污染，现有 opt-out 和 machine choice 保留。不能称与旧入口“没有告知”的默认逐字等价。前端/Lark 无新 schema、配置 owner 或控件，原 usage-settings HTTP roundtrip 已验证；完整 wheel 的原前端资源构建、校验和隔离安装也通过，但不冒充新 transport E2E 或现用 runtime 升级。

语义与 CI 对齐

原声明、source/lease fence、quota 扣额及状态 owner 不变；这是 S2 单一权威与 S10 恢复成本的增量，不是 TS cutover 或预算晋级。按当前 Goal 的 review policy，未查询或等待 GitHub CI。本 head 的十文件完整回归168项通过，无跳过；Ruff、编译、配置内 mypy19文件与 diff 检查通过。保留此前一轮未改动 usage timing 断言失败：2.817秒超过1.650+1秒；四次独立 base/candidate 重跑和之后两轮完整矩阵通过，未放宽断言。宿主负载不是已证明的失败归因。旧28项 ABBA有候选变慢样本，不能推断整批稳定提速；没有扩大20秒预算。回滚只需撤回此 bootstrap delta，不涉及状态格式迁移。

我的整体评价

APPROVE 的范围是上述完整、可逆的入口切片。长程状态与后续调用契约保持，权限拒绝与恢复通过真实路径验证；用户入口减少无关加载、保留诊断，并明确披露默认告知变化。与在旧入口底部加 wrapper 或新建 parser/权限 owner 相比，当前方案是更小的有效修复。剩余风险是共享宿主的延迟波动及尚未达成的整批20秒验收；这两者没有被文件数、启动中位数或测试通过数量冒充关闭。合并仍由维护者决定，现用安装与 provider 晋级尚未发生。

English verdict: APPROVE - exact head 440cefc; reuse the existing source/console bootstrap and Goal-scoped guard, preserving independent Goals and real scheduler recovery. 168 local tests, static checks and isolated packaged-wheel smoke passed; the frozen whole-batch20s acceptance remains open, and no live deployment is claimed.