fix(benchmark): preserve full checkpoint preparation within trial phase

[loopx-agent]

Full automatic controller checkpoints can miss mature research boundaries because the same 30-second window includes complete native/research preparation, stability checks and live-source capture. The original work correctly continues after an optional miss, but that does not establish checkpoint coverage.

Give both full controller capture hooks one shared 90-second window, capped by the original trial phase and carried unchanged into the controller. Post-release sealing retains its separate 30-second allowance and the same phase cap. Other capture declarations retain 30 seconds; direct controller calls retain their explicit timeout. First-attempt deduplication, source/cursor validation, original opportunity/cooldown, and default-off behavior are unchanged.

Validation:

• Focused capture, handoff, CLI and archive tests: 106 passed, 15 platform/explicit-integration skips.
• Isolated Linux Docker/native HTTP checks: 95 passed; one external-controller native fixture remained explicitly skipped. Separate real native-session qualifications used a mature retained point: full visible history and research evidence adopted in a non-root offline receiver; full Docker/controller/native-state recapture and byte verification passed with conserved history, incumbent, opportunity and remaining phase. No inference or grading was invoked by those real-case qualifications.
• Premerge: all direct checks and 19 selected catalog/risk/boundary checks passed. The existing maintainability ratchet remains advisory. Benchmark-sensitive manual review remains required; this is not merge readiness or merge authorization.
• Ruff, compile, diff hygiene and added-line public/private scans passed. Real-case artifacts stay private.

Placement/refactor: reuse the existing benchmark provider transport owner and one shared window constant; no new module, protocol vocabulary or parallel decision owner. Core typed quota/artifact semantics and evaluator behavior are unchanged. The bilingual provider guide discloses the explicit-on timeout change. This bounded reliability qualification does not prove sustained capture success, strategy benefit or net efficiency.

[songoow]

Folded into edgebenchtest (commit c1eb30315) as the experiment-preparation aggregate branch, so the next EdgeBench 24h run can pin one commit. Content is unchanged: 072178167 is an ancestor of that head, with the same 90-second capture window and phase cap. Leaving this PR open for its own review; no merge authority is claimed here.

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max · observation_id=b03ff20e1b97e1566d2adbf1d42f9e0fcedd546825a3c8466e93a9093f157504

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

更正原评审：撤回此前“普通文件入口仍30秒”的 P2。 我把缺少 fresh_replan 的 controller 请求直接交给 stub，绕过真实 parser，构造了生产合法入口无法到达的反例。peer 指出后已用真实文件解析与 preflight 转换复核；这是我的评审错误，不要求作者修改合法入口来满足非法测试。原失败日志保留为无效反例，不作为产品缺陷或性能证据。本完整结论替换同一条原评审，未改 PR head，也未合并。

动机

维护者需要在原始研究决策边界保存完整 checkpoint，即可供接续的代码、研究与原生会话来源。完整准备超过30秒时，原可选捕获会错过该点；本提案把显式完整 controller 捕获的观察窗口增加到90秒，同时继续服从原研究 phase 的累计截止时间。合法完成 preflight 与中断 hook 均获得这一增量，普通非 controller 路径保持30秒；没有新增人工准备或确认轮次。本次只验证可独立回滚的观察预算修复，不授予执行、评分、合并权限，也不证明科学收益、安装态或持续规模资格。成熟真实工作量的全 controller 捕获、持续运行和研究效果仍归原研究 Task；90秒只是有界观察预算修复。

改动思路

精确 head 072178167e8e3389986502922495f2cc9a32e919，实际共同祖先 b543d076995423544b0eeb1990d59d7480f48339。完整声明由原 parser 验证，controller 必须同时指定 fresh_replan 和三个 state participants。preflight 在原 typed replan 边界生成有效请求，删除原 trigger，再经 supplied_request 调用完成捕获；中断 hook 直接使用已验证请求。共享90秒常量放在现有 checkpoint provider，准入、原始来源和 writer barrier 留在原 owner。当前 PR 只交付观察预算增量，不扩大原 phase、机会、冷却或执行权限。

路径	合法行为
controller 文件缺 fresh_replan	parser 拒绝，不能成为普通捕获反例
合法 fresh 文件直接 after-hook	pre_replan_hook_required，不晚捕原决策点
preflight 有效 supplied request / 中断 hook	90秒并受原 deadline 限制
普通非 controller / 功能关闭	原30秒 / 无额外IO

这是源代码及测试路径说明，不是成熟原生性能图。增加观察预算有真实完整历史消费者，不能以裁剪历史取得较短时间；它也不能放宽 frozen trial 或原 phase 授权上限。

具体改动

完整审阅5文件 +104/-2 和两个带 DCO 的提交：共享 CHECKPOINT_CAPTURE_SECONDS=90，两处合法完整 controller 捕获消费；封存 CONTROLLER_SEAL_SECONDS=30 未变；中英文 README 披露观察预算变化；六个虚拟时钟用例保留原 phase 上限和可选失败不重放工作。

关键代码：parse_checkpoint_request 的55–58行拒绝缺失 trigger/participants；capture_before_replan 的有效请求删掉 trigger，保留原始声明 bytes 与 pending guard，才调用 capture_after_wake。因此后者按 supplied_request 判断90秒与真实入口相符。CheckpointWake._capture 按已验证的自身 request 选择同一常量，并与原 monotonic wake deadline 取最小值。controller_completion_deadline 保持独立封存窗口及原 phase。错误反例漏掉了前两个必经步骤，本次四个真实文件/preflight用例可以直接抓住该错误。

规范来源：docs/architecture/rfcs/hierarchical-agent-stride-control-v0.md#77-research-state-branching-and-matched-continuation，不可变版本 b543d076995423544b0eeb1990d59d7480f48339；criterion 7.7 Research-state branching and matched continuation，以及同版本 benchmark README 的 Controller-only native capsule transport，均按原完整来源、原决策点、原预算和未知结果恢复义务核验。关联预算增量已 implemented；不把该增量当整个研究 checkpoint acceptance 完成。

对主干的风险

初次四模块独立执行 105 passed / 28 skipped。补充真实 parser/preflight 四例在实际 base/head 各 4 passed，证明非法请求不能越过 parser，合法文件必须经 pre-hook 转换。有效内部请求使用同一35秒准备矩阵：实际 base 4 failed / 2 passed，head 6 passed；20秒 phase 始终拒绝，40秒与无额外 phase 上限的90秒路径通过。原无效 env-entry 反例及私有“修复”不再作为红绿证据。

进一步使用既有 Codex 二进制、隔离 home 和本地静态响应运行实际 source suite：57 passed。这覆盖真实原生 writer/readback、CLI settlement、SQLite/typed artifacts、source/cursor/profile 变化及重复边界；没有调用付费模型或活动 Goal。虚拟时钟不能证明成熟规模性能，Linux writer/真实 guest 与全 controller 持续资格仍按各自来源保留。

Ruff、diff、semantic advisory 后的完整 semantic smoke 通过。风险 canary 4direct和18/19selected通过；1项DSH maintainability advisory 保留，实际base→6618→6622完整报告仅归一化checkout根目录后相同（93 statements/51 decisions）。整体仍为 quality_receipt_missing，没有借作者receipt、修改阈值或声称全绿；未查询或等待CI。

我的整体评价

APPROVE，绑定上述未变 head；无当前阻塞 finding。交付判断为 justified_increment；long_horizon/user_experience 为 preserved：没有复制 authority、延期原研究 phase、裁剪见证或引入新确认轮次。成熟真实工作量的全 controller 捕获、持续运行和研究效果仍归原研究 Task；90秒只是有界观察预算修复。 原研究 owner 继续既有完整 qualification，公开批准不等于长期可靠性或研究效果成立。窗口选择已复用原 provider，有界 future-facing pass 不需再加通用超时层；应保留这次“真实 parser→preflight→有效请求”的反证资格，防止再次用不可达 stub 输入制造结论。

English verdict: APPROVE — 6618@072178167e8e3389986502922495f2cc9a32e919. Correction: the earlier file-entry P2 is withdrawn; its mocked request bypassed mandatory fresh_replan parsing and was unreachable from a legal declaration. Real file/preflight checks pass4/4 on base and head; valid35s window cases improve from expected4fail/2pass to6pass while20s phase remains binding; 57 real native/CLI/artifact source cases pass. This is a bounded observation-budget increment, not sustained controller/scientific qualification or merge authority. The independent quality receipt hold remains.