Release 1.4.0: reconcile version anchors and six-Turn catalog regression

[loopx-agent]

Prepare LoopX 1.4.0 from current main. This PR changes only release metadata and the catalog assertion missed when product and benchmark review cadence was unified at six settled work Turns. Both version owners, the generated manpage and bilingual book anchors agree at 1.4.0.

The final diff preserves main's new-Goal default: canonical SQLite with hard leases. Existing Goals retain their recorded route, and explicit File/SQLite provider and supported execution-policy choices remain available. The separate sustained/default qualification gaps remain visible; version preparation is not proof that those gates passed.

Shipped product groups in the planned tag range

• Native Chat and Lark continuity: original creation and queued Turn identity, substantive streamed answers, verified worker files returned to the original private conversation, installed transports and bounded recovery.
• Semantic control and evidence: canonical typed Todo owners, fresh quota/scheduler decisions, six-Turn receipt-backed review, scope-bound Explore capture and acceptance-bound progress reviews.
• SQLite default for new Goals and reviewed recovery: canonical creation, existing-Goal migration and cold-source import, original operation readback, later-write retention and reversible trial exit. Sustained/platform qualification is still separately tracked.
• Benchmark and cross-platform reliability: task-owned budgets, bounded best-only official feedback, native host continuation, current packaging provenance and UTF-8/Windows fixes.

Community Contributors

External community contributions are highlighted here; the project agent is credited separately. The links name representative merged work from the exact release range.

• @mikamikasuki — Runtime and installer compatibility, safe source reads, and durable regression fixtures. #5754, #5919, #6005, #6138, #6143
• @Duang777 — Goal lifecycle fencing, quota and scheduler freshness, archive recovery, PostgreSQL identity rotation, retained static rollback validation, and cadence eligibility. #5725, #5862, #5943, #6127, #6201, #6211, #6213
• @hhyykk — Canonical Todo dependencies and scheduled waits, unreadable lease protection, periodic-report recovery, and workspace task identity. #5642, #5853, #5884, #5926, #5939
• @JasonBuildAI — Windows qualification, UTF-8 CLI output, and canonical progress-review digest readback. #6073, #6118, #6119, #6200, #6204
• @songoow — Workspace typography and layout, scoped recovery, invalid ancestry enforcement, and acceptance-bound progress reviews. #5923, #6064, #6065, #6066
• @LIHUA919 — Installer review guidance, private Lark return provenance, provider failure fixtures, and test dependency consistency. #6062, #6075, #6098, #6142, #6212
• @Jim-jimu — Scoped benchmark identity, metric scale comparison, and ready-only Explore dispatch suggestions. #5758, #5887, #6022
• @GZY-SUPER-HACKER — Resolved-thread automation admission and bound-App host delivery windows. #6124, #6129
• @sakurahello1 — Native untracked source paths and typed export rejection of local references. #5896, #5897
• @Hsuehtan — First-time release contributor: shared public-safety owners for credential questions and compact opaque identifiers. #5876, #5880
• @loopx-agent — Project agent contributions: native Chat/Lark continuity, typed ownership, reviewed SQLite import/recovery, Explore evidence capture, benchmark feedback, source build recovery, and regression ownership. #5743, #5763, #5805, #5971, #6193, #6199, #6226, #6214, #6235, #6222, #6236, #6238

Compatibility and validation

Version preparation and its exact-head review are merged. The release source is now frozen at 1b452b00a114bd9f0d9c24868c9f45df42661748, including the separately reviewed runtime, packaging and regression repairs. At that source, 515 public smokes, 360 affected Python checks, static checks, a real PyPI 1.3.1-to-candidate upgrade and isolated installed UI/managed-host skill readback pass. Parent full-suite failures are retained and their affected regressions are repaired. The owner-authorized DS V4.1 Flash/high portfolio passed all 22 scenarios twice and six contrasts with 44 actor calls, zero failures/skips; prior Doubao evolving/high failures remain unqualified. The named v1.4.0 release is published at the frozen qualified source. Public tag archive and source stable installs are verified at 1.4.0. PyPI and desktop publication are queued for hosted runners; their public artifact readback is pending and desktop-stable remains on 1.3.1. Post-freeze test improvements are excluded from this release range. External backend/platform and sustained SQLite qualification limits remain explicit.

中文产品变化

• 原生 Chat／Lark 保留原创建与排队 Turn 身份、实质答案、来源绑定的文件回传及 transport 恢复。
• canonical 类型化 Todo、quota／调度新鲜度、六个结算工作 Turn 的复核、作用域绑定的 Explore 证据及验收绑定的进展复核。
• 新 Goal 保留主线 canonical SQLite／hard lease 默认；已有 Goal 可审阅迁移、冷来源导入和保留新增写入的可逆退出，长期资格单独记录。
• benchmark 任务预算、best-only 官方反馈、Host 延续、打包来源及跨平台修复。

社区贡献者

以下突出外部社区贡献，项目 Agent 单独署名。链接列出本次真实发布范围内的代表性合并工作。

• @mikamikasuki — 运行时及安装兼容、安全来源读取与回归夹具。 #5754, #5919, #6005, #6138, #6143
• @Duang777 — Goal 生命周期隔离、quota 与调度新鲜度、归档恢复、PostgreSQL 身份轮转、保留静态文件回滚验证与周期资格。 #5725, #5862, #5943, #6127, #6201, #6211, #6213
• @hhyykk — canonical Todo 依赖与定时等待、不可读租约保护、周期报告恢复及工作区任务身份。 #5642, #5853, #5884, #5926, #5939
• @JasonBuildAI — Windows 资格验证、CLI UTF-8 输出和 canonical 进展复核摘要回读。 #6073, #6118, #6119, #6200, #6204
• @songoow — 工作区排版、作用域内恢复、无效 ancestry 拒绝和绑定验收的进展复核。 #5923, #6064, #6065, #6066
• @LIHUA919 — 安装器评审说明、私有 Lark 回传来源、provider 失败夹具和测试依赖一致性。 #6062, #6075, #6098, #6142, #6212
• @Jim-jimu — benchmark 作用域身份、指标尺度比较和仅对可执行 Todo 给出的 Explore 建议。 #5758, #5887, #6022
• @GZY-SUPER-HACKER — 解析目标线程后才安装自动化，以及绑定 App 的 Host 交付窗口。 #6124, #6129
• @sakurahello1 — 原生未跟踪来源路径与类型化导出中的本地引用拒绝。 #5896, #5897
• @Hsuehtan — 首次进入正式发布的贡献者：统一凭据判断和短 opaque id 的 public-safety owner。 #5876, #5880
• @loopx-agent — 项目 Agent 贡献：原生 Chat/Lark 连续性、类型化归属、可审阅的 SQLite 导入恢复、Explore 证据、benchmark 反馈、源码构建恢复与测试归属修复。 #5743, #5763, #5805, #5971, #6193, #6199, #6226, #6214, #6235, #6222, #6236, #6238

发布边界

发布源码冻结为 1b452b00a114bd9f0d9c24868c9f45df42661748。515 项公开 smoke、360 项相关 Python 回归、静态检查、真实 PyPI 1.3.1 升级到候选包、隔离安装版 UI 与托管 Host 技能回读通过。父提交完整测试的失败记录保留，相关断言已修复并在冻结源码回归。经授权替代的 DS V4.1 Flash/high 完整资格通过：22 个场景各两次、六个对照、44 次 actor 调用、零失败与跳过；先前 Doubao evolving/high 失败保留且未通过。v1.4.0 release 已从冻结合格源码发布，公开 tag archive 与源码 stable 安装回读通过，源码 stable 已为 1.4.0；PyPI 和桌面发布仍等待 runner，制品回读待完成，desktop-stable 仍为 1.3.1。冻结后测试改造不在该版本范围；外部后端、未测平台与 SQLite 持续资格不因版本准备而宣告通过。

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; runtime_reported; reasoning_effort=xhigh

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

精确 head c91602d06ff8d13fca6815f70eb5da20008b38cd：无阻塞发现；此结论批准版本准备，未认证最终发布或长期 SQLite 资格。

动机

发布维护者和核对安装版本的用户。准备 1.4.0 时，源码、打包、手册和书籍仍标作 1.3.1；配置目录测试还期待已被主线改为六个工作 Turn 的旧五-Turn 默认。候选的两个版本源和当前手册统一为 1.4.0，旧断言与独立的六-Turn 主线规则一致。不更改新 Goal 的 SQLite 默认、不迁移已有 Goal、不调整权限，也不以版本准备替代最终发布资格。完整发布验证、发布物升级回读及个人图文指南仍待最终发布源码验证；SQLite 持续资格缺项继续保留。

改动思路

复用两个既有版本源、原生成器和已有配置目录测试，避免增加版本同步机制或再次改动已合并的存储默认。本 PR 仅交付八个文件的版本元数据与一个已有测试断言。实际调用链仍是 CLI 读取 loopx.__version__，打包器读取 project metadata，安装新鲜度和升级计划使用既有 release manifest，手册由原生成器产出。职责是版本投影，不新增存储策略、状态机或 provider IO；故障与修复继续由现有版本一致性检查及安装新鲜度归属处理。最小改动不能只改一个版本源，但也无需增加同步框架。原始 SQLite 回退已完全撤出最终 diff。

具体改动

规范依据 docs/product/release-readiness.md，固定 revision 449cd8bf51b7eeac6f00646a1ca03b37d3f4c7bc。Version metadata：implemented，两个版本源、man 和四个中英书籍锚点统一为 1.4.0，版本、手册和书籍验证覆盖。Compatibility Gate：deferred，最终 merged source 的完整 lane、发布物升级和指南回读仍待执行，PR 不发布 tag 或 stable。

关键代码讲解

• loopx/__init__.py:5 的 __version__ 从 1.3.1 改为 1.4.0，未改 version 命令的 JSON 结构；现有 CLI／manifest 消费此值。
• pyproject.toml:7 的 [project].version 与运行版本一致，继续由既有打包流程产生 wheel／sdist metadata；无需新版本 reader。
• man/loopx.1:1 改生成标题；四个 book 文件只替换当前版本锚点，未改历史发布记录或操作义务。
• test_inspection_lists_registered_namespaces_without_local_refs 将旧五-Turn 断言改为已合并规则的六-Turn，未放松命名空间、隐私或配置内容断言。原规则由既有 goal vision policy 决定，测试未从实现动态获取预期值。

对主干的风险

最大风险是版本标签一致却误把候选当作已经发布，或只改运行版本导致包 metadata 和安装判断不同步。现有版本 smoke 同时检查两个源、manifest、update plan，并用 9.9.9 错配验证必须进入 repair，而不能显示 fresh。精确整合 head 的 101 项存储／配置回归通过，涵盖默认 SQLite、显式开启／关闭、原操作重试与后续写入；36 项 cadence/default 验证补足旧断言来源。生产存储、租约、quota／调度和权限代码与 base 完全相同，不增加用户导航或干预。Ruff 的仓库正式 scope、mypy 19 个文件、TypeScript、版本／man／book／update 和 integrated risk canary 均通过；未查询 CI。宽泛 ruff check . 曾扫描无关 benchmark/deprecate 文件失败，这不是规范要求的 lint lane，不将其改称通过。

我的整体评价

long_horizon 为 preserved：此 PR 不改变后续 Turn、重启、回执或副作用路径。user_experience 为 improved：同一候选在运行、打包和文档中版本一致，现有错配修复仍有效。元数据更新和一个 durable test 的修复相称，代码量没有增长；未来边界检查确认现有 owner 足够，无需新增抽象。语义词汇未新增，default_off 与 authority 语义在此净 diff 不适用。独立的发布及 SQLite 持续资格继续保留 passed／failed／untested 区分，不能由此 review 授予发布默认资格；若最终源码或声明变化，应重验对应 lane。

English verdict: APPROVE - c91602d; coherent 1.4.0 metadata and six-Turn catalog oracle, preserving main SQLite defaults. Exact-head local version, mismatch/update, 101 focused Python, static, semantic, man/book and risk-canary checks passed. Final release and sustained SQLite qualification remain separate.