feat(stride): add read-only shadow stride observation (M1)

[huangruiteng]

Summary

• Add loopx/control_plane/runtime/stride_observation.py: a read-only, shadow-only projection that derives a provider-neutral hierarchical_stride_observation_v0 from existing public-safe run receipts.
• Delivery layer: material slices, latest outcome, evidence freshness (6h window).
• Authority layer: bounded slices since the last replan/vision/gate run; segment disposition stays unknown.
• Effect layer: explicit unknown: True (fail-closed; host metrics absent).
• Offline evaluator reports only derivable mismatch signals (settlement_lag) and recommends nothing.
• 4 focused tests with public-safe fixtures.

Issue Or Task

• Tracks [Research]: qualify hierarchical stride control across effect, delivery, and authority boundaries #3203 (RFC docs(rfc): define hierarchical agent stride control #3204 M1 read-only observation; merged as ae31e498)

Validation

• python -m pytest -q tests/control_plane/test_stride_observation.py — 4 passed
• ruff check clean on both files
• python -m py_compile passes

Type of Change

• New feature (read-only observation; no runtime behavior change)
• Test update

Boundary Checklist

• No credentials, private state, local paths, or raw evidence included.

[huangruiteng]

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

详细中文评审：只读 Shadow Stride Observation M1（PR #3207）

评审目标 head：a0f81edb5aeeb9395bafe742ce07d2c2b5535f77（base: main，2 文件，+214/-0；Tracks #3203，RFC #3204 M1）

内容与边界

• 新增 loopx/control_plane/runtime/stride_observation.py：从现有 public-safe run receipt（runs/index.jsonl）派生 hierarchical_stride_observation_v0，只读、shadow-only（shadow_only=True 恒真），不触碰 Todo 选择、scheduler、quota、通知、gate 或执行路径。
• delivery 层：material slices、latest outcome、evidence freshness（6h 窗口）；authority 层：自最近 replan/vision/gate run 以来的 bounded slices，segment disposition 保持 unknown；effect 层显式 unknown: True（fail-closed，符合 RFC §8.1/§11“缺失 host 指标不得猜测”）。
• 离线 evaluator 只报告可派生 mismatch signal（settlement_lag），recommendations 恒为空。
• 4 个聚焦测试：空 runtime fail-closed、字段派生、陈旧证据信号、跨 agent 归因隔离。

与 RFC #3204 的一致性

严格落在 M1 最小切片内：只读 observation + offline evaluator + public-safe fixtures；无新 capability、无 executor、无 adaptive policy、无 shadow recommendation 之外的任何动作。放置于 control_plane/runtime（最近的 status read model 区域），符合 scope-fit。

非阻塞问题

无。合入前等 CI（pytest）完成即可；后续 M2（effect-stride 实验）应另开 todo/PR。

我的整体评价

窄小、诚实（未知即未知）、可逆，且用测试钉死归因与 fail-closed 语义。结论：APPROVE。

---

English Verdict (PR #3207)

• Exact head reviewed: a0f81edb5aeeb9395bafe742ce07d2c2b5535f77
• Verdict: APPROVE (author-owned PR; GitHub blocks formal self-approval, recorded as COMMENTED)
• Key findings: Adds a read-only, shadow-only hierarchical_stride_observation_v0 projection over existing public-safe run receipts (delivery material slices/evidence freshness, authority bounded-slices-since-change, explicit unknown effect layer) plus an offline evaluator that reports only derivable signals (settlement_lag) and recommends nothing. No runtime, quota, scheduler, notification, gate, or execution behavior changes; 4 focused tests cover fail-closed empty state, field derivation, stale-evidence signals, and cross-agent attribution isolation; ruff and py_compile are clean.

[huangruiteng]

CI status: pytest / build / dependency-review all SUCCESS at head a0f81ed. Merge-ready pending owner merge.