[loopx-agent]

Compare staged replan context on the refreshed experimental baseline. The treatment is exactly the reviewed #7197 diff, with no manual conflict resolution or additional implementation changes. Its patch SHA-256 matches the source PR byte-for-byte.

Both arms share the current-main integration baseline from the preceding proposal, checkpoint state, runner settings and native opportunity history. #7196 is excluded. This proposal records source integration only; models have not been launched and behavioral benefit remains unqualified.

Validation on the integrated treatment:

• 100 checkpoint, retry and replan Python tests passed; 4 platform-specific skips.
• 63 TypeScript replan/stride tests and the control-plane typecheck passed.
• Bilingual RFC and public-boundary checks passed.
• Two isolated network-disabled Linux native Codex checkpoint forks preserved the captured full-history digests, with no credentials or model calls. This verifies history restoration, not execution admission or model benefit.
• Standard premerge initially failed missing local TypeScript dependencies and an existing scheduler-route fixture. Dependencies were installed and the semantic check passed in both arms. The scheduler-route fixture independently fails in the common baseline with canonical_authority_required before replan projection. This remains a disclosed failure, not a passing check.

在最新公共实验基线上集成 #7197，补丁与原 PR 完全一致，两臂仅此差异，未包含 #7196。验证覆盖恢复、重试、事实读取和复盘写回；真实原生历史恢复通过。标准检查的依赖缺失已恢复，基线同现的 scheduler-route fixture 失败仍保留。尚未启动模型实验，不宣称行为收益，也不授权合并 main 或更新正在运行的实验。