test(stride): add synthetic stride-boundary shadow fixture

[NIU-123370]

Adds one synthetic stride-boundary fixture for the read-only stride shadow observation (M1 #3207, RFC #3204), following the standing CONTRIBUTOR_TASKS.md invitation ("Add one synthetic lifecycle or stride-boundary fixture").

Test-only change: +241 lines appended to tests/control_plane/test_stride_observation.py under a single banner comment; zero runtime changes (git diff -- loopx/ is empty), so the canary baseline is untouched.

What the fixture proves

Four boundary tests, each pinned to an RFC §13 validation criterion (the mapping is cited in the test comments):

• test_boundary_observation_stays_shadow_only_across_segment_states — criterion 8: shadow_only=True, effect.unknown=True, and empty recommendations across four synthetic segment states, plus a byte snapshot of the runtime tree before/after every build+evaluate proving the projection never mutates what it observes.
• test_boundary_missing_metrics_stay_unknown_not_inferred_from_prose — criterion 2: when classification/delivery outcome are missing and marker words appear only in unread summary/notes prose, the projection reports material_slices=0, latest_outcome="none", segment_disposition="unknown" — never substring inference.
• test_boundary_authority_changes_require_explicit_markers — criterion 6: status/monitor/review-style classifications never restart authority-window counting; only explicit replan/vision/gate markers do, and the latest marker wins.
• test_boundary_projection_replays_byte_identical — criterion 3: two builds across two distinct runtime roots with identical synthetic receipts collapse to one unique sorted-key JSON serialization.

Fixture hygiene (per the invitation's constraints)

• Explicit activation: hand-built synthetic run-index receipts only — no source/draft/review bodies, provider payloads, private locators, cursor state, or apply/publish authority.
• No local paths or credentials anywhere in the fixtures.
• Deterministic by construction: freshness offsets pinned at 45m and 12h, keeping replay assertions away from the 6h evidence-freshness wall-clock boundary.

Tests

python3.11 -m pytest tests/control_plane/test_stride_observation.py -q → 8 passed (4 pre-existing + 4 new), 0.03s.

Tracks #3203.

[huangruiteng]

Review: test(stride): add synthetic stride-boundary shadow fixture

Reviewed exact head: 321bcbae734dfde15645e570c75891e99d7fd6a7 (test/stride-boundary-synthetic-fixture), base main.

动机

RFC #3204 的 M1 stride shadow 观察目前只有 4 个基础测试，覆盖空目录、常规派生、陈旧证据与跨 agent 归属，但对 RFC §13 的边界准则（shadow-only 不变量、prose 不得被推断为指标、显式 authority marker 才重启计数、确定性 replay）没有回归锚点。本次按 CONTRIBUTOR_TASKS.md 的邀请新增一个合成边界 fixture，把四个最关键的只读投影契约钉死，防止后续实现 drift。

改动思路

纯测试改动：在 tests/control_plane/test_stride_observation.py 末尾追加 241 行，全部基于手工构造的 synthetic run-index receipts（复用已有 _write_run_index/_run 辅助函数），每个测试直接断言 build_stride_observation/evaluate_stride_observation 的输出。没有复制任何 provider payload、私有 locator、host 路径、凭据或 cursor 状态，canary baseline 未触碰（git diff -- loopx/ 为空）。

具体改动

四个测试分别对应 RFC §13 的一个准则：

• test_boundary_observation_stays_shadow_only_across_segment_states（准则 8）：四种 segment 状态下 shadow_only=True、effect.unknown=True、recommendations=[]，并用构建前后的字节级快照证明投影从不改写被观察的 runtime 树。
• test_boundary_missing_metrics_stay_unknown_not_inferred_from_prose（准则 2）：classification/delivery_outcome 缺失、而 "replan/gate/progress" 等词只出现在 summary/notes 时，material_slices=0、latest_outcome="none"、segment_disposition="unknown"——与模块只读 classification 与 delivery_outcome、从不读 prose 字段的实现一致。
• test_boundary_authority_changes_require_explicit_markers（准则 6）：只有含 replan/vision/gate 的分类重启 bounded_slices_since_change 计数，status/monitor/review 分类不重启，且最新 marker 生效（marker 在 index 4 时计数为 1；无 marker 时 3 条全计）。
• test_boundary_projection_replays_byte_identical（准则 3）：相同 receipts 在两个不同 runtime root 各构建两次，序列化结果唯一——因为观察负载不包含 runtime_root 路径，确定性成立。

对主干的风险

• 无运行时改动，风险面仅限测试文件本身；误判只会让 CI 捕获契约回归，不会改变任何 shipped 行为。
• 时序确定性：freshness 偏移为 45m/5h/12h，5h 距 6h 新鲜边界仍有 1h 余量，测试在毫秒级内完成，不存在跨边界翻转的实际风险；datetime.now(UTC) 相对偏移是现有测试文件的既有模式。
• 体积方面 4 个测试 +241 行偏厚，但每个都对应独立 RFC 准则且复用现有 helper；若后续补齐更多准则，可考虑参数化共享骨架，非本次阻塞项。
• fixture 卫生已扫描：无凭据、私有路径或内部链接，git diff --check 干净。

验证

• 本地：tests/control_plane/test_stride_observation.py 8 passed（4 旧 + 4 新）。
• CI：Sign-off / dependency-review / pytest 全部成功，无 pending。
• 语义核对：逐条对照 loopx/control_plane/runtime/stride_observation.py 实现（marker 子串匹配、latest-marker 覆盖、delivery_outcome=="outcome_progress" 计数、输出不含 runtime_root），断言与实现一致。
• 变更仅 1 个文件、+241/−0，单 commit 且含 Signed-off-by。

整体评价

这是一个高质量、边界明确的合成 fixture：每个断言都能回溯到 RFC §13 的验收准则，且验证的是可复用的只读控制面契约，符合 durable smoke 保留标准。无阻塞项。

---

English verdict: APPROVE — test-only synthetic stride-boundary fixture with four assertions mapped to RFC #3204 §13 criteria (shadow-only read contract, no prose inference, explicit authority markers with latest-wins, byte-identical replay); verified against the implementation, no runtime changes, fixture hygiene clean, local 8/8 and all CI checks green. Reviewed at exact head 321bcbae734dfde15645e570c75891e99d7fd6a7. No blockers; optional P2 is to parameterize the shared state-dict skeleton if more §13 criteria follow.