[loopx-agent]

Change

The base/head CLI output smoke already freezes one caller workload for both arms. Load the copied fixture support before the fixture imports it, so an older checkout can run the current observation helpers. Each subprocess still imports the product CLI and budget owner from its own selected checkout.

Validation

• The new subprocess regression reproduces an import error with the old loading order and passes with the fix. The main branch proposal passes 198 focused tests and Ruff.
• Combining this patch with the independent Stride example fix passes the complete local CLI output budget smoke. The strict comparison ran 98 base and 98 head rows with zero review signals. The standalone base still exceeds its existing crowded Turn plan growth allowance by 32 characters. Stride fix #8325
• The same loader reached the real C1 CLI and exposed existing size and output-shape gaps that the import error had masked. The C1 proposal retains the budget and reports the observed failures. C1 #8266

The standalone Standard canary reports the inherited Turn plan budget failure and an outstanding exact-commit change-quality receipt.

[loopx-agent]

Reviewer: model_agent, gpt-6.1-sol, OpenAI, runtime_reported, reasoning_effort=max

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

对 00fc80f 给出独立通过结论。没有阻塞发现。

动机 / Motivation

维护 CLI 输出预算的人需要在两个 checkout 中运行同一份测量夹具。 当前 workload 需要新增 helper，旧 checkout 只有旧 helper。旧顺序抛出 ImportError，新顺序进入测量。 main 与候选各98项，两个完整receipt逐项相等，比较器为零差异和零review signals。 当前提案修复跨版本验证器。Stride 输出压缩和 C1 切换由原提案验收。

改动思路 / Approach

沿原 runner 先以 tests.control_plane.cli_output_fixture 绑定冻结 helper，再加载 workload。产品与预算继续由子进程自己的 checkout 提供。完整 diff 限于两条加载语句和 subprocess 回归，原比较器与阈值继续复用。

具体改动 / Changes

规范为 docs/development/good-smokes.md，修订 8a4d4f9。Start With The Contract 由两个真实产品的测量覆盖，Consolidate Without Losing Coverage 由原比较器、预算断言和既有测试入口承担。规范

• 旧runner在相同反例中抛出缺helper的ImportError，候选通过。
• probe runner与differential两模块共198项通过，Ruff通过。
• main与候选各98行，完整receipt逐项相等，比较器为零差异和零review signals。

对主干的风险 / Risks

独立预算smoke仍触发拥挤计划的11,432 > 11,400，main的原生矩阵也复现同一失败。产品和上限在本diff中逐字未改。Quick的5项直接检查通过，4项所选检查中保留该预算失败，精确提交质量回执仍缺失。Stride提案

我的整体评价 / Verdict

通过当前局部harness修复。它让跨版本比较到达真实产品输出，原增长拒绝继续生效。渲染组件没有变动，本轮覆盖源码测量和本地回归。Windows安装、模型采用和完整成本未资格验证，独立合入门继续按原owner执行。

English verdict: APPROVE. The frozen helper loads before its consumer on 00fc80f. Both real checkouts emit identical 98-row receipts. The inherited Turn plan growth failure remains visible.