fix(turn-driver): import the Turn result owner from its generated module

[songoow]

Summary

• M2 (feat(semantics): generate shared Turn contracts and controller rules (M2) #4499) moved LoopXTurnResultKind into turn_contract_generated.py and left transaction.py as a compatibility re-export. The Python producer scanner (loopx/semantics/python_production.py, _qualified_bindings) attributes a production site only when the owner class is imported from the owner's own module, so all LoopXTurnResultKind.* production sites in executor.py became unknown_producer after M2 even though executor.py did not change.
• Import the enum from the generated owner (one line). It is the same enum object through both paths; behavior is unchanged and the compatibility export stays for external readers.
• Other transaction importers (command_validation.py, managed_step.py, settlement.py, __init__.py) only read or compare the enum and have no production sites, so they are left alone to keep the diff to the measured effect.

Issue Or Task

• Refs [Task]: Semantic vocabulary convergence follow-ups after RFC PR #4433 / 语义词表收敛后续任务 #4447 — Track B, B2 producer pilot for the four Turn kernel vocabularies (turn_result_kind). Not a closer for the tracker.
• Contributor task ID: n/a

Validation

• Tested revision: cda1133
• Run state: finished
• Input classes: public_fixture

Check kind	Result	Public-safe evidence / limitation
static	passed	python -m py_compile loopx/control_plane/turn_driver/executor.py; import identity asserted: transaction.LoopXTurnResultKind is turn_contract_generated.LoopXTurnResultKind is executor.LoopXTurnResultKind.
regression_parity	passed	python examples/semantic-vocabulary-drift-smoke.py --report on main vs head: unresolved_producer_sites 52 → 42, distinct sites 34 → 32. Resolved: executor.py::_host_result_stage (2), _run_task_validator (7), _task_validation_stage (1). No site becomes newly unresolved; every other summary line identical; smoke still ok.
integration	passed	loopx canary premerge --from-git-diff --git-diff-base origin/main (tier standard, run from the checkout): ok: true, 10 checks selected, 0 failures.
unit	not_run	pytest is not installable in the authoring sandbox; CI test-shard is the authoritative run. No test patches executor.LoopXTurnResultKind by path, and no anchor pins the unresolved-site count (PRODUCER_VOCABULARY_ANCHOR / RETURN_PRODUCER_ANCHOR cover registry coverage, not this metric).

• Coverage and gaps: the change is an import-source swap for one symbol; the parity row measures the only intended effect (producer attribution) and shows no collateral change. The remaining 42 unresolved sites are pre-existing dynamic/parameter paths (e.g. project_turn_route, _ControllerInputs.render, driver.py::build_loopx_turn_plan) that need the B2 slot analysis, not an import fix.

Frontend / Visual Evidence

N/A — no user-visible change.

🤖 Generated with Claude Code

[huangruiteng]

动机

评审 head：cda113317f1416df6b13420c1ce06b537c20649f；base：main（merge-base f1166e81e）。

这是一条被测量出来的回归修复：#4499(M2) 把 LoopXTurnResultKind 搬进生成模块 turn_contract_generated.py，并把 transaction.py 留成兼容 re-export；而 Python producer 扫描器（loopx/semantics/python_production.py::_qualified_bindings）只在"owner 类从 owner 自己的模块被导入"时才承认一个生产点。于是 executor.py 里所有 LoopXTurnResultKind.* 写入点在自己没改动的情况下变成了 unknown_producer。

代价是证据层面的：M2 的目的正是让"Turn 结果词表确实被生产"这件事可被证明，而这次 attribution 丢失恰好削弱的就是这个不变量；运行期行为没有任何变化。

更小的修法只有这一条：要么回退 M2（把三套词表 owner 重新拆开），要么让扫描器学会穿透 re-export 链（语义更大、且带来误归因风险），要么把每个 importer 都改一遍（未被测量的改动）。作者选了"只改有生产点的那个模块"的一行，并明确把其余 importer 留给 Track B B2 分析——这个取舍是对的。

改动思路

入口是 examples/semantic-vocabulary-drift-smoke.py --report，权威输入是注册表派生的 owner 表（loopx/semantics/vocabulary_v0.json）加上各模块真实的 import 语句，判定权在 scan_python_production/_qualified_bindings。

改动本身是一次 import 源切换：LoopXTurnResultKind 不再从 .transaction（兼容 re-export）取，而是从 .turn_contract_generated（owner）取。运行期绑定的是同一个类对象（我在进程内断言过 transaction.LoopXTurnResultKind is turn_contract_generated.LoopXTurnResultKind is executor.LoopXTurnResultKind），所以行为不变，变的只是扫描器能否把这个模块的生产点归到 owner 名下。

这个"import 源具有语义"的规则本身没有被本 PR 修改，也没有为了这一处去特化扫描器——这符合仓库"每个词表一个 owner、兼容层要有理由"的既有约定。

具体改动

1 个文件、+1/-1：loopx/control_plane/turn_driver/executor.py 的 import 块——从 .transaction 的导入组里删掉 LoopXTurnResultKind，改为单独一行 from .turn_contract_generated import LoopXTurnResultKind。executor 内 30 处枚举使用点全部保持不变。

关键代码讲解

executor.py:76 的这行 import 是唯一的语义位点：扫描器对 ImportFrom 的别名做 (_module(path), name) 查表，只有解析到 owner 模块才把 owner 类放进 bindings；解析不到就把它从 bindings 移除，后续该模块的写入点被报成 unknown_producer（而不是猜一个 owner）。

python_production.py:141 _qualified_bindings 是这条规则的实现者，本 PR 没有改它——这正是这条修复值得肯定的地方：作者没有为了让自己的模块通过而去放宽规则，而是把自己的模块挪回规则之内。我在同一个 worktree 里把 import 改回 .transaction 复现了基线（52），改回 owner 后回到 42，且 unknown 列表的差异恰好是 PR 声称的 10 条（_host_result_stage 2 条、_run_task_validator 7 条、_task_validation_stage 1 条），没有任何新增行。

对主干的风险

最强回归场景是"以后有人再次把 import 挪到 re-export 上"，此时 attribution 会静默退化：unresolved_producer_sites 从 42 涨回 52，但 smoke 仍然 ok、215 条聚焦测试也照常通过——也就是说这条修复目前没有门禁保护。这是下面 P3 的依据，而不是反对本次合并的理由。

爆炸半径只在语义清单的证据层：没有运行期、状态、CLI 或权限面改动。我实测：head 上 smoke ok（unresolved_producer_sites=42）；回退 import 后 ok（52）；两边都没有新增 unresolved 行；pytest 聚焦 215 例通过（含 tests/test_loopx_turn_transaction.py、tests/test_loop_turn_loop_controller.py、tests/architecture/test_semantic_production.py、tests/architecture/test_turn_contract_generation.py）。测试里没有按路径 patch executor.LoopXTurnResultKind，所以这次 import 源变更不会打断任何 mock。

P3（非阻塞）：把这条被修好的指标钉住。建议给 unresolved_producer_sites（或至少 executor 这几行）加一个锚点断言，让"再次回退 import"变成红灯，而不是只让一个数字变化。

P3（非阻塞）：PR 描述里"The other transaction importers … have no production sites"这句偏强。loop_controller.py 仍有 7 处枚举成员使用，其 _ControllerInputs.render/_envelope_route 仍在 unresolved 列表里（属于本 PR 明确延后的动态/路由型路径）。范围判断没问题，但措辞会让下一位读者误以为那些模块已被归因；建议改成"这些模块没有新增可归因点，其既有动态路径保持 unresolved"。

我的整体评价

baseline（f1166e81e）与 head 的对比：unresolved_producer_sites 52→42，unknown 列表只有删除、无新增，枚举 identity 与运行期行为不变，smoke 两边都 ok。这是一条可复现、可回滚、作用范围精确的修复。

体量与影响匹配（change_proportionality: proportionate）：一行修复一个被测量的证据回归，没有新增抽象、没有放宽扫描器规则、没有碰兼容导出。repository_reuse: reused（复用既有 owner 与兼容层），typed_state_rule 保持"未解析即 unknown"的失败方式，authority_semantics 不适用（无 actor/权限面）。

结论 APPROVE，两条 P3 均非阻塞。复评只需在 head 变化时重跑上述 smoke 两条状态对比与聚焦测试。

English verdict: APPROVE - exact head cda1133; the one-line import move restores producer attribution for executor.py without changing runtime behavior or the compatibility re-export. Reproduced both directions in one worktree: unresolved_producer_sites 52 -> 42, the unknown-list diff is exactly the ten claimed executor sites with no additions, enum identity holds through both paths, smoke ok and 215 focused tests pass. Two non-blocking P3s: the repaired metric is not pinned by any anchor (a revert stays green), and the PR body's claim that the other importers have no production sites is imprecise for loop_controller.py.