fix(validation): align prompt budgets and Goal source recovery

[loopx-agent]

Problem and resulting behavior

Cumulative Goal-outcome instructions outgrew older heartbeat presentation budgets, while one Goal test still imposed 3,300 characters despite the existing 4,000-character host contract. A recovery fixture also expected successful admission after deleting the only Goal source. This proposal retains all generated instructions and the missing-source refusal, updates the measured presentation budgets, and verifies refusal followed by recovery against actual File/SQLite authority.

The scheduler contract smoke now checks the shared timestamp parser used by the current production hint transport, replacing a removed private helper while retaining its identity and UTC-normalization assertions.

Author declaration

Written by: model agent, Codex / OpenAI, using the LoopX Agent repository identity. Base: main at b9e60c59f3c367a63755d97f4595d9c2eef91947. Reviewed proposal head: 1ec369aa4199129957eebb5726f34291ed4bab03.

Acceptance	Implementation and evidence
Preserve complete Goal requirements, authority and continuation instructions	24 immutable-base/head generated-packet comparisons: bodies, command routes and every other field are equal; only presentation max_chars / within_budget differ. Includes memory disabled/enabled and zero/320-character profile scopes.
Keep bounded presentation diagnostics	Full/compact/brief base ceilings change from 12,000/7,200/4,800 to 12,800/7,600/5,400. Existing 640-character opt-in memory headroom is unchanged. Exact-limit and one-character-over cases still pass/refuse. Thin, JSON-envelope and native Goal limits are unchanged.
Retain fail-closed source loss and canonical recovery	All File/SQLite open/done and missing/stale cases retain one fallback snapshot and unchanged canonical rows; missing Goal refuses work, restoring its original bytes recovers.
Retain shared scheduler time semantics	Full monitor scheduler contract smoke plus 240 scheduler tests pass, including real native projection and timezone-equivalent inputs.

Validation and limits

• Seven complete related Python modules: 360 passed, zero skipped; the original two failing modules account for 46 of those passing cases.
• Original focused baseline: 9 failed / 37 passed. The earlier complete-suite attempt stopped at maxfail with 2 failed / 9,006 passed / 603 skipped / 1,903 passed subtests; remaining selected tests were unexecuted. These results stay failed/incomplete.
• A required premerge smoke independently failed on immutable base at the removed _parse_monitor_timestamp reference. The repaired complete smoke passes; no baseline waiver or skip is used.
• Ruff, compile/diff hygiene, staged bilingual check and bounded public/private scan passed. Diff-scoped vocabulary advisory found no supported new carriers; its coverage limits remain.
• Standard premerge rerun passed:6 direct and17 selected checks,zero failures/warnings/manual holds. Exact-head review/readiness remain separate before merge. The broader full-suite qualification remains separate; this proposal does not claim whole-tree health.

The budget decision evidence records the matched pre-outcome/current measurements and disclosed headroom. No prompt prose is shortened to obtain a green budget. Generated instructions add no content in this repair. No execution, spending, scheduling, store, provider activation or release authority changes; no paid/model calls or installed-adoption claim.

UI impact: none. Existing CLI diagnostics and validation contracts only; no new capability or frontend journey.

中文：保留完整成果验收与继续执行指令，按相同负载修正展示预算；缺失 Goal 源码仍拒绝执行，并验证恢复与 canonical 状态不变。旧 scheduler smoke 改读现有共享时间解析器，保留 UTC 语义。360 项相关测试、完整 smoke 与逐字段对比通过；原全量失败和未覆盖范围保留，整体健康尚未验收。

[loopx-agent]

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

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; runtime_reported; effort=max.

动机

维护者在验证宿主 Goal 提示及丢失展示文件的恢复路径时会遇到这组失败。原先完整提示被旧展示预算判为超限，恢复测试又要求删除唯一 Goal 源码后仍准许执行；现在保留全部指令，并验证拒绝及恢复。生成入口按实测负载报告预算，真实 File/SQLite 读回仍使用同一快照且不改 canonical 任务。本 PR 不启用记忆、不授予执行或费用权限，也不证明全仓健康。

改动思路

展示预算继续由现有 heartbeat owner 计算，恢复和时间语义沿原生产 owner 验证；无需新增状态、兼容分支或提示框架。当前 PR 完成已复现的提示预算与陈旧测试修复；全仓剩余覆盖、安装采用和模型实际遵从仍分别验收。不修改提示正文来追逐字符阈值；保留原始成果、必读上下文、权限、结算与停止条款。24组不可变base/head完整packet对比覆盖4种Goal宿主、4种heartbeat模式、memory关闭/开启和0/320字符profile，仅展示max/within字段发生已披露变化。缺Goal的原生拒绝继续保留，恢复原文件后重新读回同一File/SQLite权威状态，无额外确认或安装步骤。

具体改动

spec_ref: AGENTS.md；spec_revision: b9e60c5。采用变更前的Goal导向与预算决策规则，新增测量表不是自授验收依据。

• measure-same-contract：同一输入比较累计成果指引前后增加523字符；当前full/compact/brief关闭memory、320 profile的归一化正文为12688/7531/5285。基准输入和原失败均保留。
• preserve-meaning：现有预算owner仍只计算展示诊断；完整正文、命令及其他字段逐值相同。memory固定640余量和Goal宿主4000硬边界不变，缺失源也未获得执行权。
• document-budget-decision：三个base上限12000/7200/4800调整为12800/7600/5400，分别保留112/69/115字符的实测最大负载余量；thin、信封及执行费用边界不变。所有模式仍核验exact-limit和one-over。
• goal-source-boundary：File/SQLite完整恢复测试保留单snapshot、归档依赖及canonical不变断言，先要求缺源拒绝，再恢复原bytes重新验证健康。既有scheduler smoke只改用当前生产导入的共享parser，并保留精确UTC及新增identity断言，不恢复已移除的私有API。

七个完整相关模块360项通过、无跳过；完整scheduler smoke通过；标准premerge为6direct及17selected全过，无失败、warning或manual hold；CQR精确scope有效。原两完整模块9败/37过与更早全量2败/9006过/603跳过/1903子例通过仍保留为失败和不完整；不是把未执行部分拼成通过。

对主干的风险

预算元数据对full/compact/brief的memory关闭与开启路径均有明确数值变化，这是展示回归预算修正，不是整体packet完全不变或执行准入放宽。原生Goal硬限、source-health、scheduler/quota/default activation均未改变。尚未执行的全仓部分、已安装采用和模型实际遵从未验收，不能从本次360项或当前PR批准外推。该变更无需持久化迁移，可独立回滚三个常量及对应验证；CI按当前策略未查询。

我的整体评价

未发现阻塞问题，APPROVE。此PR完成可独立核验的原生诊断与验证修复，long_horizon保留、user_experience改善：完整要求与停止语义不丢失，缺源不错误放行，恢复回到健康状态，零新增用户交互轮次。展示预算继续由现有 heartbeat owner 计算，恢复和时间语义沿原生产 owner 验证；无需新增状态、兼容分支或提示框架。当前 PR 完成已复现的提示预算与陈旧测试修复；全仓剩余覆盖、安装采用和模型实际遵从仍分别验收。本文是作者自审的COMMENTED批准结论，正式合并仍分别受原生readiness与已有维护者授权约束。

English verdict: APPROVE. The existing owners and all prompt instructions remain; measured presentation caps and stale validation expectations are corrected. Native source-loss refusal/recovery and shared timestamp semantics pass. Whole-tree health, model adoption and installation remain separate qualifications.