docs(runtime): ground context-cost qualification in current CLI evidence

[loopx-agent]

Required-work context now preserves fresh empty-hook observations and later action-specific reads, but the CLI budget fixtures already exceeded their ceilings. The bilingual M7.4 convergence checkpoint now records matched immutable-base/current-source stdout costs, identifies useful context versus separately reviewable diagnostics, and gives the bounded next implementation a concrete acceptance. It also replaces stale Chinese wording about pending reads with the current owner and fulfilled-source boundary.

This is a characterization prerequisite for the existing heartbeat/TurnEnvelope convergence work. It changes no budget, runtime, provider, execution authority or transport default, and does not claim installed-host or model-quality qualification.

Validation:

• Compared aa87cc019 and fc411c878 with the identical public fixture bytes and argv, retaining 12 matched stdout rows per checkout. Crowded Turn: 16,115 → 17,403 characters; thin heartbeat: unchanged at 3,122. Compact JSON and emitted stdout are reported separately.
• The original probe-runner/differential selection remains 3 failed / 1 passed on both revisions: crowded Turn and the enabled multi-subagent envelope exceed their original ceilings; actual stdout-growth rejection passes. The nested runner path workload is explicitly distinguished from the table. No historical failure is relabeled as a pass.
• Current real File/SQLite heartbeat registry-route, bound-wait recovery and required-source checks: 18 passed, zero skips.
• Source-checkout premerge: 11 selected checks + 3 direct checks passed, no warnings or manual holds; public boundary scan clean. The initial semantic smoke lacked npm dev dependencies; npm ci --ignore-scripts and the source-checkout rerun resolved that environment failure.
• Semantic-inventory advisory: no changed source vocabulary; git diff --check clean.

The future-facing pass updates the existing convergence checkpoint and its Chinese mirror rather than adding a new roadmap or runtime abstraction. The next implementation must audit diagnostic consumers, preserve full routes and freshness semantics, and jointly repair the existing budget owner and runner tests. SQLite release-default, installed App adoption and sustained quality/cost acceptance remain separate.

[loopx-agent]

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

Exact head: d5bd985fa20044dc12bddf4393012bac95bda06f
Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; xhigh

动机

维护者验证运行前必须读取的任务上下文时，需要判断预算失败来自真实增长还是测试路径差异。
原 runner 在历史基线也超限；本次固定源码、公共负载与路径后记录基线和当前输出，区分新增执行事实、诊断和排版开销。
维护者现在能定位具体成本和真实未通过的上限，并按既有 owner 的有界任务继续实现。
本 PR 不修复预算或改变上限、默认值、运行时和权限，也不证明安装态宿主、模型质量或 SQLite 发布默认已完成。

仍待完成既有 M7.4 上下文形态资格：诊断消费者核验、无损压缩或合理余量判断，以及原 runner 与语义差分修复；安装态 App 和模型质量另行验收。

这是既有 M7.4 收敛工作的有用前置切片，不把记录数字当作预算修复。下一步已明确为 runtime/quota owner 对诊断消费者、无损压缩与合理余量的联合预算/runner 实现；安装态 App、模型质量及完整默认资格继续单独验收。

改动思路

按 docs/development/testing-and-quality.md，固定版本 fc411c8786be89a193220896124a06582a297d8b 的 Budget Failure Decisions 与 Roadmap-Aligned Optimization 逐项核验：同负载计量、信息价值、原失败与硬约束保留、下一有界结果均已实现为本次 checkpoint。保持旧上限和只提高上限都不能代替判断；当前方案先把真实来源说明白，没有引入新框架或平行 authority。

具体改动

完整审查两份 RFC：英文增加同 fixture/argv 的字符与紧凑 JSON 对照及下一实现验收；中文镜像这些数值，并替换过时的读取描述，明确完整当前任务来源已交付、混合 Goal 仍需全文读取、后续 action 仍要新鲜来源。crowded Turn 的 484 个 envelope 紧凑字符和 510 个诊断紧凑字符有独立归因，诊断的副作用披露职责保留给实际消费者审计。

对主干的风险

无运行时、权限、默认值、预算或格式改动；两文档 diff 可整体回滚。最强反例是把表格误当通过：原 runner 同基线和当前版本仍 3失败/1通过，crowded Turn16115→17403超过16000，嵌套路径下 enabled envelope11291→11961超过9000，真实 stdout 膨胀拒绝仍通过。文档明确该路径负载与表中不同，并禁止据此宣称模型或安装态资格。实际12对stdout的fixture字节和argv相等；真实File/SQLite路由/等待恢复/完整来源18通过0跳过。source premerge11项及3个直接检查全通过，公共边界干净；初次缺npm开发依赖的环境失败经安装依赖与source重跑恢复，未改门限。

我的整体评价

APPROVE：本 PR 交付可归因的成本 checkpoint 与一致的中英文当前事实，正好支持已有有界下一实现，不伪称修复预算或整体收敛。future-facing检查已采用替换过时镜像与复用现有checkpoint；无生产结构可进一步删除。未查询或等待CI，未授权合并。

English verdict: APPROVE - d5bd985: useful, reproducible bilingual cost characterization; original budget failures remain open, while source premerge and public-safety checks pass. No runtime, threshold or default change.