fix(turn): fence journal commits by GoalRef

[Duang777]

Summary

• classify Turn journals as legacy, exact GoalRef, or invalid without changing loopx_turn_journal_v0
• hand the source alias guard from Python to the TypeScript journal owner so every source journal mutation validates the current exact GoalRef and writes atomically
• preserve legacy RPC shape and bytes, add ABA/replay/expired-handoff coverage, and qualify only the turn_journal M3 inventory row

Verification

• npm run -s typecheck:control-plane -- --pretty false
• npm run -s test:control-plane (3,514 passed, 30 skipped)
• focused TypeScript journal suite (17 passed)
• focused Python Turn/journal/Host suites (97 passed and 50 passed)
• architecture inventory and registry census (10 passed)
• docs-governance-smoke.py and repository-hygiene-smoke.py
• loopx canary premerge --from-git-diff (11 selected, 0 failures)

Mypy isolated output remains identical to the baseline: 28 existing errors in the same three touched modules, with no new-line diagnostics.

Scope

This qualifies only the Turn-journal owner. It does not remove the overall M3 activation hold or enable execution authority. PR #5278 remains open and overlaps the bilingual RFC/inventory, so those metadata edits may need conflict resolution if it lands first.

Refs #5206

[Duang777]

CI follow-up after merging current main@649826221 into this branch:

• 25 checks passed, including all TypeScript shards, Stage 2C, Windows, dashboard acceptance, Ark 3.11/3.13, DCO, and the release build.
• Frontstage Pages / build fails on the existing v1.2.3 Developer Book release-anchor mismatch. The same failure occurs on main run https://github.com/loopx-project/loopx/actions/runs/36610043069, and docs: synchronize 1.2.3 Developer Book release anchors #5318 contains the dedicated fix.
• Python shards 2/3 report test_new_independent_twin_cannot_hide_behind_generated_pair plus both test_live_decision_adds_only_existing_required_read_channel cases. All three failures reproduce unchanged in a clean detached worktree at origin/main@649826221; none of their test or implementation files differ in this PR.

I have not folded these unrelated baseline repairs into #5324. The branch is clean and mergeable; review is still required. No merge was attempted.

[huangruiteng]

结论：APPROVE。评审完整 base-to-head diff，exact head 9b3873f7e64b28d6c39871de19836d05e614d8d2；本次 journal 提交隔离已独立验证，没有发现阻塞问题。这是 #5206 的 M3 journal 增量，不是整个 source-session profile 已可启用的结论。

动机

同一个 Goal alias 被替换为新实例后，旧 Turn 的结果不应继续推进旧 journal，更不应拿旧身份结算。只在 Host 启动或返回时检查一次不够：验证、恢复和后续 checkpoint 期间仍可能切换实例。本 PR 将保护落实到每次实际 journal 写入，解决可复现的 A→B 期间旧 A 继续提交问题。

改动思路

复用既有 source alias guard 和 TypeScript require_current 决策，Python 只在持有 guard 时传递当前 authority 与锁见证，TS 写入 owner 验证后再获得 journal lock、执行原单调状态转换和原子写入。guard 持续覆盖提交，不把 plan 中的 GoalRef、锁 token 或 profile 名称当成新授权。全部 checkpoint 归入同一 persist_journal 回调，减少原先分散的特殊包装；legacy journal 保持原 wire 与只读历史检查能力。

具体改动

关键代码讲解

• parseTurnJournalGoalBinding：将无引用、完整 exact 引用、不完整/不一致引用分为 typed legacy/exact/invalid。plan 和 transaction 的两份引用必须一致，不用文本包含规则推断实例身份。
• source_journal_admission：在现有跨运行时 alias guard 中读取规范 authority，并交出一次内部锁见证。它不写入 journal、不授予 peer/session 权限，非 source 路径返回原 legacy 行为。
• commitTurnJournal：exact journal 必须提供匹配的 admission、registry/GoalRef 和有效活锁见证，再调用已有 require_current。随后沿用 journal 的不可变 plan、checkpoint 顺序和原子提交规则；旧 A 或错误见证不能改写 A/B journal。
• persist_journal：初始提交、Host 尝试、验证、失败、恢复和结算 checkpoint 共用此路径。source 拒绝转换为现有 runtime rejection，避免只保护 happy path 或另建 Python 决策源。

本人用同一脚本在 merge-base 649826221289cd4cb3dd8880d016e0afbbaca0fc 与本 head 执行实际 Turn executor→managed TS RPC→File journal/跨运行时锁。6 个场景涵盖 legacy/current-source 完成与重跑、Host 前/过程中/验证时换实例、source 缺少引用；完整归一化结果除目标修复场景外保持一致。关键反例中 base 在验证后错误提交旧 A 并结算；head 拒绝 stale A，原 journal 字节不变、旧 A 无 writeback/spend/scheduler，新 B 随后成功提交。Host 与业务结算回调使用合成实现；journal 提交和锁不是 mock，因此证据证明提交边界而非真实模型或外部业务效果。

对主干的风险

新增严格条件只服务已有 source-session 精确绑定：缺失/错误 GoalRef 或失效 guard 会拒绝；legacy 的完整执行、持久 JSON 与重复调用均做了 base/head 对照，未被迫升级。运行时 RPC 是同包边界，持久 loopx_turn_journal_v0 仍须保留兼容；新 admission 是瞬时提交输入，不能充当可重放授权。没有新增用户确认步骤、CLI 开关或 UI 编辑器；历史 inspect/恢复入口通过 companion 测试。

语义与 CI 对齐

本人本地验证：171 项 Python、48 项 TS、typecheck 通过，canary 全部选定检查及 5 项直接检查通过。包括真实 File/managed TS admission、单调 checkpoint、不可变 plan、失败恢复、历史读取和 CLI companion 路径。相关 #5278 的 handoff owner 已合并；本 PR 没有借此宣称其他 inventory 行或完整 M3/M4/M5 都已完成。未拉取或等待远端 CI；也未改预算、放松既有检查。此次不涉及 PostgreSQL authority store 重构，未以 mock 替代受影响的真实 journal 后端。

我的整体评价

增量边界合理：关闭一个完整的 journal commit/recovery 缺口，保留现有 RFC/issue 中整体激活仍待验证的阶段。未来维护性检查已应用在统一提交回调和复用 typed owner 上，下一次 checkpoint 规则修改无需再寻找分散的 admission 包装；更大的 executor 重构不属于这次必要范围。拒绝旧 A 之后已验证新 B 恢复有用推进，不只停留在 blocker 回执。批准本 exact head 的增量，整体 profile 启用及合并由各自 owner 判断，不自合并控制面代码。

English verdict: APPROVE — exact head 9b3873f. Real managed-TS/File journal counterexamples verify per-commit GoalRef fencing, legacy parity and recovery to a fresh instance. This qualifies the bounded M3 journal slice, not overall source-session activation.