refactor(status): single-source the source-registry shadow findings vocabulary

[YZJF]

Refs #4447 (Track A) — checklist row: "Classify the 19 historical multi-value twins and 18 semantic string forks; retain justified independent contracts."

This is one bounded sub-item, not completion of the row and not completion of the tracker. It must not close #4447.

Classification: SOURCE_REGISTRY_SHADOW_FINDINGS is same_semantics

The name has four definitions on main (a28562e97):

Location	Values	Verdict
loopx/status.py:289	{source_registry_missing, stale_source_registry}	active — merged
loopx/control_plane/status/registry_health_projection.py:28	same	active — retained as owner
deprecate/benchmark-legacy/.../benchmark_status_runner.py:218	same	outside active scope, untouched
deprecate/benchmark-legacy/.../benchmark_status_compaction.py:185	same	outside active scope, untouched

Evidence that the active pair is one concept rather than two contracts:

• Same values. Both sets are exactly {source_registry_missing, stale_source_registry}.
• Same slot. Both are passed as the source_registry_shadow_findings keyword argument.
• Same consumer. Both reach work_items/attention_queue.py:77, which is the only place the value is read (if finding.get("kind") in source_registry_shadow_findings and live_items).
• Same runtime, no persisted boundary. The set is an in-process argument, not a serialized field, so there is no historical-reader compatibility argument for keeping two copies.

Per the issue's own rule, only confirmed same_semantics groups become merge work — this one qualifies, so it is merged.

Change

registry_health_projection.py keeps the only definition; loopx/status.py imports and re-exports it.

• The dependency direction stays outer → control_plane. status.py already imports .control_plane.status.collection and .control_plane.status.runtime_summaries at module level, and registry_health_projection.py has no import back into status.py, so no cycle is introduced.
• status.SOURCE_REGISTRY_SHADOW_FINDINGS is kept as a public name on purpose: examples/control_plane/attention-queue-readmodel-smoke.py:62,95 reads it as status_module.SOURCE_REGISTRY_SHADOW_FINDINGS. Removing the re-export would be an unforced break of an existing caller, which the row's "caller parity" requirement argues against.
• Values are unchanged, so no behavior changes.

Verification

• New regression tests/control_plane/test_source_registry_shadow_findings.py locks three things: the admitted values are unchanged, loopx.status exposes the same object as the projection rather than a copy, and no literal can be reintroduced into loopx.status.
• examples/control_plane/attention-queue-readmodel-smoke.py → ok (this is the caller that reads the re-exported name).
• ruff check clean on both changed files.
• tests/control_plane -k "status or registry or attention or queue" — no new failures versus baseline.

Out of scope

The remaining 18 semantic forks and 19 multi-value twins in that row are not classified here; each needs its own caller check. The two deprecate/benchmark-legacy/ copies are left alone. Base a28562e97, Python 3.13.12.

[huangruiteng]

动机

#4447 Track A 把 SOURCE_REGISTRY_SHADOW_FINDINGS 判为同名同值的 fork：loopx/status.py 与 loopx/control_plane/status/registry_health_projection.py 各自持有一份内容完全相同的两值集合，而它们喂给的是 attention-queue read model 的同一个 source_registry_shadow_findings 参数。两份字面量只要有一处以后被改动，读模型接受的 finding 就会和生产者发出的种类不一致。

改动思路

既然语义相同、且投影本来就是该 read model 输入词汇的 owner，就把定义收在投影里（registry_health_projection.py:28 保持不变），loopx/status.py 改为 re-export 同一个对象，保证既有调用方（examples/control_plane/attention-queue-readmodel-smoke.py 通过 status_module.SOURCE_REGISTRY_SHADOW_FINDINGS 取值）不需要改动。新增 3 个用例分别钉住取值集合、对象同一性（is）以及"status 里不得再出现第二份定义"。

具体改动

• loopx/status.py（+6/-4）：删除本地字面量，改为从投影导入并在原位置继续使用（第 16 行导入、第 1093 行传入 read model），并加注释说明这是 re-export 而不是第二份定义。
• tests/control_plane/test_source_registry_shadow_findings.py（新增 42 行）：test_vocabulary_is_unchanged、test_status_shares_the_single_definition（is 同一性）、test_status_defines_no_second_copy（源码里不得再出现 stale_source_registry）。

关键代码讲解

1. loopx/control_plane/status/registry_health_projection.py:28 — SOURCE_REGISTRY_SHADOW_FINDINGS = ('source_registry_missing', 'stale_source_registry')：现在是全仓唯一的活动定义，并且在同文件第 53 行直接喂给 read model，因此"owner 与消费点在同一模块"这件事成立。
2. loopx/status.py:16 — from .control_plane.status.registry_health_projection import SOURCE_REGISTRY_SHADOW_FINDINGS：re-export 是这次改动的兼容性关键。我确认了消费点仍在 loopx/status.py:1093，且 shipped smoke examples/control_plane/attention-queue-readmodel-smoke.py:62/:95 仍以 status_module.SOURCE_REGISTRY_SHADOW_FINDINGS 读取——如果直接删掉名字而不 re-export，这个 smoke 会立刻失败。
3. tests/control_plane/test_source_registry_shadow_findings.py:35 — status.SOURCE_REGISTRY_SHADOW_FINDINGS is registry_health_projection.SOURCE_REGISTRY_SHADOW_FINDINGS：用对象同一性而不是值相等来防 fork，这一点比"值仍然相等"强得多，因为未来的漂移往往从"又是一份字面量"开始。
4. tests/control_plane/test_source_registry_shadow_findings.py:41 — assert "stale_source_registry" not in inspect.getsource(status)：这条最直接，但也是我唯一提出的问题（见下）：它检查的是源码文本，而不是赋值结构。

对主干的风险

没有阻塞项。 这是一次同语义去重：取值集合未变（('source_registry_missing', 'stale_source_registry')），loopx/status.py:1093 与 read-model smoke 的消费路径都保留；我全仓搜索确认活动代码里只剩投影这一处声明（其余命中是 deprecate/benchmark-legacy/ 下的归档副本、global_registry_health.py 里作为生产者发出的 kind、以及 example smoke 里作为数据使用的字符串，都不构成第二份词汇表）。分支与当前 main 合并干净。

P3（非阻断，断言形式）：test_status_defines_no_second_copy 断言的是 "stale_source_registry" not in inspect.getsource(status)。这意味着以后有人在 loopx/status.py 的注释或 docstring 里提到这个取值（例如解释 re-export 的来源），测试会因为散文而失败，而不是因为真的出现了第二份定义——这类假阳性会诱使后来者删注释而不是复核 fork。最小修复是改成检查解析后的模块级赋值目标（或把检查范围限定在赋值语句），保留"重新引入字面量就失败"的能力。

验证（全部在 be53db7d9 上跑）：pytest -q tests/control_plane/test_source_registry_shadow_findings.py → 3 passed；python examples/control_plane/attention-queue-readmodel-smoke.py → attention-queue-readmodel-smoke ok；git merge-tree --write-tree HEAD origin/main → 无冲突。按本 lane 配置不拉取 CI。

我的整体评价

APPROVE。这是"删除重复权威"这一类里最干净的一种：owner 已经在正确的模块（投影就是 read model 输入的来源），改动只是把另一份相同字面量换成 re-export，并用对象同一性测试把 fork 的可能性真正关掉；同时它照顾到了既有消费方，因此没有连带 churn。我有意把它与"值相等"型断言区分开评价：is 断言是这次最值钱的三行。唯一的 P3 是那条源码文本断言的形式，属于可随手收紧的细节，不影响正确性。

English verdict: APPROVE — exact head be53db7d9ffe1313e8999e2aeb2fb0db0d1b24d0 of #4517. The duplicate SOURCE_REGISTRY_SHADOW_FINDINGS vocabulary is merged into its existing owner (loopx/control_plane/status/registry_health_projection.py:28) and loopx/status.py now re-exports the same object, so the attention-queue read model keeps receiving identical findings while a second literal can no longer drift. Validation at this head: the 3 new tests pass (value set, is identity, no re-introduced literal), the shipped consumer smoke examples/control_plane/attention-queue-readmodel-smoke.py prints ok, and the branch merges cleanly with current main. A tree-wide search shows no remaining active declaration outside the projection. One non-blocking P3: the anti-fork assertion inspects module source text, so a future comment mentioning the finding name would fail it without a real re-fork.