feat(manager): type declared evidence-source health in the turn window

[huangruiteng]

The manager turn context declares every evidence source, but only with source_id, status and scope. A source that was declared and read nothing therefore had no typed representation: the window could say declared_unread_sources: ["ssh:ark-devbox"] while nothing in the packet said what that means for the answer, so staleness could only be narrated as a caveat, and a stale or unread source was indistinguishable from a source that reported no progress.

This adds one typed health row per declared source, reusing the vocabulary the read refusals already use (source id, typed reason, coverage effect, next action) and the remote read path's next-action wording:

{"source_id": "ssh:ark-devbox", "source_host": "ark-devbox", "status": "not_read",
 "freshness": "stale", "reason": "declared_source_not_read_this_turn",
 "coverage_effect": "no evidence was read from this source; the answer must not present it as no progress",
 "next_action": "Read this source with the manager evidence read tool, or let the next Turn's source rotation dial it."}

• freshness is a closed set (current | stale | unknown) declared as MANAGER_SOURCE_FRESHNESS_VALUES.
• A source read in this window is current with no reason, no coverage effect and no next action.
• The local source is typed the same way: reachable but unread in this window is unknown with local_evidence_not_read_this_turn, so a window that read nothing cannot be answered as if it had.
• not_configured keeps the declaration's own reason (ssh_alias_not_configured) and gets the register-and-retry next action.

The change is additive on the packet: sources, declared_unread_sources, read_status and every other existing field keep their exact shape, so no current reader changes behaviour. Owning boundary is the manager evidence window in loopx/chat_manager_context.py; no new capability, provider, state or CLI surface is added.

Entry points: the manager/agent answer path consumes the packet directly, so no frontend or Lark companion change is needed; the packet is what the steward prompt already receives. Verified by reading the manager turn-context consumers in this repository.

Validation:

• uv run --extra test python -m pytest tests/test_chat_manager_context.py -q -> 24 passed (3 new cases).
• uv run --extra test python -m pytest tests/test_manager_ssh_evidence.py tests/test_chat_project_coordination.py tests/test_chat_manager_inspection.py tests/test_manager_context_tracking.py -q -> 52 passed.
• uv run --extra test python -m pytest tests -q -k 'manager or steward or chat or coordination' -> 938 passed, 9879 deselected.
• uv run --extra test ruff check loopx/chat_manager_context.py tests/test_chat_manager_context.py -> all checks passed.
• The new assertions cover the row builder for a read window, an unread local window, a declared-not-read remote source and a missing-alias source, plus one turn-context assertion that the health rows cover every declared source in declaration order.

Successor slice of the evidence-source-health row (steward RFC M2 anchor A11); the remaining live re-measurement of raw provider text in answers stays open on the parent row.

Control-plane change (loopx/**): proposed for review and left for the maintainer to merge.

[huangruiteng]

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

精确 head：6af5edb639478350a5c4b8c655f9bbaed631fd6f（base be7789fd9）。无阻断发现，建议合并；合并决定留给 maintainer。

动机

管家每轮的 evidence window 会声明每个证据源，但只用 source_id / status / scope 表达。所以「声明了却没读到」这件事在包体里没有类型化表示：declared_unread_sources: ["ssh:ark-devbox"] 只是把 id 列出来，没说什么；local 源无论这一窗口读没读，永远是 available。结果就是模型只有两条路：要么省略这个缺口，要么把它写成「数据可能过期」这类免责声明——而读者既看不出是哪个源没读到、覆盖了什么，也分不清「没读到」和「这个 Goal 没有进展」。

改动思路

给每个被声明的源加一条类型化的 health row，复用读失败（#4787）与远端读路径已有的四元组：源 id、类型化原因、覆盖率影响、下一步动作；并复用 ssh 路径的 source_freshness 词表，把新鲜度收敛成闭集 current | stale | unknown。渲染放在同一个 owner 内、由同一个 sources 列表派生，因此 source_health 与 declared_unread_sources 不可能互相矛盾。

具体改动

• loopx/chat_manager_context.py：新增 MANAGER_SOURCE_FRESHNESS_VALUES 等 10 个常量（含那句关键的覆盖率影响文案）、纯函数 _source_health_rows(sources, read_status=...)，以及 _evidence_window 返回体里新增的 source_health 成员。
• 分类规则：available 且本窗口读过 → current，reason/coverage_effect/next_action 全为 null；available 但没读 → unknown + local_evidence_not_read_this_turn（窗口读了个空，不能被当成读到了）；not_read → stale + declared_source_not_read_this_turn；not_configured → unknown + 声明自带原因 + 注册别名并重试；任何未来状态落到 unknown + source_status_<status>，不会静默变成 current。
• tests/test_chat_manager_context.py：新增 3 个用例（读窗口与未读窗口的 row builder、远端源与缺别名源的类型化、turn context 里每个声明源都有一行且顺序一致）。

对主干的风险

1. 纯增字段：sources、declared_unread_sources、read_status、窗口边界与其它成员一律不变，唯一新增的是 source_health；938 个 manager/steward/chat/coordination 测试保持全绿。
2. 对现有读者的影响：只有对 evidence_window 做整字典严格比较的读者才会观察到差异，仓库内的断言都按具名成员读取。
3. 反证：把这个 head 的测试文件放到 base be7789fd9 上跑，两个新用例都以 KeyError: 'source_health' 失败，说明断言确实在测新增契约而不是现有输出。
4. 诚实边界：这次改的是「包体里有类型化事实」，没有强制答案文本引用它们；父任务的「真实答案里原始 provider 文本的复测」仍然是验收项，未被这次改动关闭。

我的整体评价

正向且比例合适。它把一句本来只能靠模型自觉写的免责声明，换成包体里可被断言、可被引用的一条类型化事实，且没有引入新模块、能力、状态或 CLI 面；分类规则集中在一个纯函数里，词表与既有的读失败/远端源保持一致，非法状态（未来的新 status）也不会静默降级成 current。没有阻断问题。

English verdict: APPROVE - the manager evidence window now types every declared source's health (freshness, reason, coverage effect, next action) instead of leaving a declared-but-unread source untyped; the change is additive on an existing packet with one pure helper and three new cases, verified by a base/head counterfactual (KeyError: source_health on base), 24 passed in the changed suite, 52 in the adjacent suites and 938 in the manager/steward/chat/coordination selection with ruff clean. It does not yet force the answer text to cite the rows, so the parent row's live re-measurement stays open. Merging is a maintainer decision.