test(rectification): prove memoization in-process; tolerate golden float drift (BUG-733)

Replace whole-structure golden == with same-process cached vs None fallback equality, plus discrete-strict / measured-tolerance golden comparison. Do not rebuild the golden JSON.

docs/BUG_HISTORY.md

@@ -11396,3 +11396,19 @@
 - 相关记录：BUG-464
 - 复发自：无
 - 修复版本：待发布
+
+## BUG-733 | 校正记忆化 golden 对全精度浮点做整体 `==`，跨机门禁靠运气绿
+
+- 状态：resolved
+- 首次发现：2026-09-15
+- 最近更新：2026-09-16
+- 影响面：`tests/test_rectification_engine_memoization.py`、`tests/golden/rectification_engine_memoization_v1.json`（文件未改）
+- 用户现象：用户看不见。门禁 glob `tests/test_rectification_*.py` 在 `CORE_PYTEST_TARGETS` 里，换一台机器或基础镜像会因 `score` / `margin_percent` 尾数差 1.0e-4～1.1e-3 而红，CI 绿只是碰巧和生成 golden 的那台舍入一致。
+- 触发条件：`test_score_candidates_matches_baseline_golden` 把 live payload 与 golden 做整体 `==`。
+- 根因：golden 存了全精度浮点并整体相等。这是 BUG-712 的同一形状。BUG-712 的防复发只落在 `ephemeris_events` 那一处（量化到 6 位再比），没有仓库级守卫，BUG-721 新写的 rectification golden 重蹈覆辙。
+- 修复：只改测试，不动 `scripts/`。主证据改成同进程差分：同一批 static contexts，A 带四层缓存、B 把 `ashtakavarga_result` / `shadbala_result` / `vimshottari_timeline` / `narayana_periods` 置 `None`，`compute_event_candidate_rows` 输出严格相等。golden 改成分档比较：离散字段严格相等；浮点用实测最大漂移 1.1e-3 推出的容差（绝对 2e-3 与相对 5e-4 取更宽者）。禁止调用 `write_golden()` 更新那份 JSON。
+- 验证：同进程 A/B 逐字相等，且 B 的 `calc_shadbala` 次数显著高于 A（本机 6 vs 0）。反向：把某一层缓存换成全零 shadbala 对象后不再相等。golden 分档比较通过；把 golden 浮点改 1e-2 必须红；把离散字段 `time` 改掉必须红。`pytest tests/test_rectification_engine_memoization.py` 与 `pytest tests/test_rectification_*.py` 0 failed。golden JSON `git diff` 无改动。
+- 防复发：任何 golden 比较都不得对浮点做整体 `==`；浮点必须量化或带容差，容差数值要有实测依据并写在注释里；能在同进程内做差分证明的命题，不得用跨机 golden 代替。
+- 相关记录：BUG-712、BUG-721
+- 复发自：BUG-712
+- 修复版本：待发布

docs/tasks/PROGRESS-rectification-engine-memoization-fix-20260915.md

@@ -0,0 +1,64 @@
+# PROGRESS · 等价 golden 跨机不稳（2026-09-15 / 执行 2026-09-16）
+
+- 执行分支：`codex/rectification-engine-memoization-fix-20260915`
+- 工作树：`.worktrees/rectification-engine-memoization-fix-20260915`
+- 任务书：`docs/tasks/TASK-rectification-engine-memoization-fix-20260915.md`
+- 开工 `HEAD` = `origin/staging` = `51a65d92`（用户核对时是 `7227b1ed`，其后合入了 BUG-732）
+- BUG：`BUG-733`（复发自 BUG-712；落库时最大号 732）
+
+## 开工前置
+
+| 项 | 结果 |
+| --- | --- |
+| `git status -sb` 第一行 | `## codex/rectification-engine-memoization-fix-20260915...origin/staging` |
+| `HEAD` | `51a65d92` |
+| BUG_HISTORY | 最大号 **BUG-732**。本单预占 **BUG-733**。读完 BUG-712 / BUG-721 |
+| 本机 Python | 主仓 `.venv` 是 25 字节指针文件。本 worktree 用 Anaconda 3.11.7 `--system-site-packages` 建真实 venv，复用 `swisseph 20230604`，未升级依赖 |
+| 先复现 golden | `test_score_candidates_matches_baseline_golden` **本机绿**。live 与 golden 逐字相同：`score` 8.6274 / 8.6227 / 8.1701，`margin_percent` 5.3006。验收机上的 1.1e-3 漂移在这台 Windows / Anaconda 3.11.7 上复现不出来——golden 就是同类机器生成的。仍按任务书改比较方式，不调用 `write_golden()` |
+
+未改 `docs/tasks/README.md` 状态列。未推 staging。未改任何 `scripts/` 实现。未改 `tests/golden/rectification_engine_memoization_v1.json`。
+
+## 做了什么
+
+| 条 | 做法 |
+| --- | --- |
+| 5.1 | 新增 `test_cached_static_context_matches_uncached_fallback`：同一请求、同一进程，A=带四层缓存的 static contexts，B=四键置 `None`。`compute_event_candidate_rows` 输出 `==`。调用计数：A 的 `calc_shadbala` = 0，B = 6（3 候选 × 2 事件）。反向：`test_poisoned_static_cache_diverges_from_live_rows` 把第一候选的 `shadbala_result` 换成全零对象，分数从 12.5125 变成 12.6125，断言不再相等 |
+| 5.2 | `test_score_candidates_matches_baseline_golden` 改成分档比较。整数 / 布尔 / 字符串 / 列表身份严格相等；浮点容差 `max(2e-3, 5e-4 * \|expected\|)`，注释写明来自实测最大漂移 1.1e-3。反向：`score += 1e-2` 红；`time = "99:99"` 红 |
+| 5.3 | 未调用 `write_golden()`。`git diff` 里该 JSON 无改动 |
+| 5.4 | 六份检查结论见下表。同类问题只记录不修 |
+| 5.5 | `docs/BUG_HISTORY.md` 新增 BUG-733，复发自 BUG-712。证据是 5.1 同进程差分 + 两次反向验证，不是「现在绿了」 |
+
+## 5.4 `tests/golden/` 检查
+
+| 文件 | 全精度浮点整体 `==` | 在哪个测试里比 | 怎么比 |
+| --- | --- | --- | --- |
+| `consultation_contract_keypaths_v1.json` | 无 | `tests/test_consultation_contract_golden.py` | 只存 keypath 与 JSON 类型（`kind=keypath_types_only`）。值不入库。live 是 keypath 超集 + 类型相同 |
+| `ephemeris_events_raman_20260915_90d.json` | 无（BUG-712 已改） | `tests/test_ephemeris_events.py` | `longitude` / `speed_longitude` 量化到 6 位再 `==`；`kind` / `date` / `body` / 星座字段严格相等。文件里仍存全精度，但测试不再整体 `==` |
+| `golden_cases.json` | 无 | `tests/test_cli_smoke.py::test_full_reading_golden_cases_cover_user_ready_output` → `tests/run_golden_cases.py` | 路径非空、模块计数、整数不变量 `ashtakavarga_sav_total == 337`、日期覆盖。不是把整份引擎浮点输出整体 `==` |
+| `qizheng_stem_branch_19900409.json` | 无（比较已量化） | `tests/test_qizheng_chart_engine.py` | 离散字段（宫、宿、ketuMode）严格相等；`siderealLon` / `mansionDegree` 用 `round(..., 2)`。**文件本身仍存全精度浮点**（如 `177.94225865978694`）。若有人新写 `assert result == golden` 会变成 BUG-712 同类。本单只记录不修 |
+| `rectification_engine_memoization_v1.json` | **本单之前有** | `tests/test_rectification_engine_memoization.py::test_score_candidates_matches_baseline_golden` | 改前：`candidate_scores` 与 `decision_receipt` 整体 `==`。改后：离散严格、浮点容差。文件未重建 |
+| `upstream_sync2/` | 无整体浮点 `==` | 见分项 | `relationship_einstein.json`：`test_upstream_sync2_goldens.py` 只比键与类型。`consultation_marriage_evidence_snapshot.json`：shape。`consultation_marriage_1990_fictional.json`：`test_relationship_event_class_evidence.py` 比离散 Dasha 字符串。`pl9_time_system_table.baseline.md`：`test_full_reading_conditional_dashas.py` 比对 markdown 表行文本（前 5 行 `==`），不是 JSON 浮点。`pl9_time_system_table.md` 与 `pl9_export_timing.json`（含 `seconds: 8.647`）**没有任何测试读取** |
+
+同类观察（不修）：qizheng golden 文件仍存全精度；ephemeris golden 文件仍存全精度（测试已量化）；`upstream_sync2/pl9_export_timing.json` 是无引用的浮点秒数。
+
+## 测试
+
+| 套件 | 结果 |
+| --- | --- |
+| `tests/test_rectification_engine_memoization.py` | **14 passed**（原 10 + 新增 4） |
+| `tests/test_rectification_*.py` | **186 passed, 0 failed**（14.42s） |
+
+命令使用 `.\.venv\Scripts\python.exe`（任务书写的 `.venv/bin/python` 在 Windows 不存在）。
+
+未改任何既有断言的期望值。新增 4 条：同进程差分、缓存毒化反向、浮点 1e-2 反向、离散字段反向。原 golden 断言从整体 `==` 改成分档，不是放宽离散字段。未跑 `run_quality_gate.py --profile quick`：该 profile 在 pytest 之后会跑 `npm test`，本机 Windows 缺 `tsx` PATH（与 BUG-721 进度记录同一缺口），本单未改前端。
+
+## CHANGELOG
+
+未改。纯测试比较方式，用户可见打分不变。
+
+## 偏离
+
+- 任务书写 `.venv/bin/python`，本机换成 `.\.venv\Scripts\python.exe`。
+- 本机未能复现验收机上的 golden 红灯（舍入与生成机一致）。仍改比较方式，并保留验收机测到的 1.1e-3 作为容差依据。
+- 未 `git push origin HEAD:staging`。
+- 5.4 按让步只检查并记录，未加仓库级自动扫描守卫。

tests/test_rectification_engine_memoization.py

@@ -2,17 +2,25 @@
 
 Golden payload in tests/golden/rectification_engine_memoization_v1.json was
 produced from origin/staging @ a8d29d1b before any memoization landed.
+
+Do not compare that payload with a whole-structure ``==``. Cross-machine
+libm / pyswisseph rounding already drifted ``margin_percent`` by 1.1e-3
+(TASK-rectification-engine-memoization-fix-20260915). Equivalence of the
+four cached layers is proven in-process instead.
 """
 
 from __future__ import annotations
 
 import inspect
 import json
+import math
 from datetime import date, datetime
 from pathlib import Path
 from typing import Any
 from unittest.mock import patch
 
+import pytest
+
 from scripts.active_rectification_event_engine import (
     _candidate_datetimes,
     _controlled_transit_rules,
@@ -35,6 +43,20 @@ GOLDEN_PATH = ROOT / "tests" / "golden" / "rectification_engine_memoization_v1.j
 FROZEN_TODAY = date(2026, 9, 16)
 TIMING_KEYS = frozenset({"column_compare_ms"})
 SOURCE_COMMIT = "a8d29d1b6cc37ff865ddec6c8bccdf9aa889ee53"
+CACHE_LAYER_KEYS = (
+    "ashtakavarga_result",
+    "shadbala_result",
+    "vimshottari_timeline",
+    "narayana_periods",
+)
+# TASK-rectification-engine-memoization-fix-20260915 §2 measured max
+# cross-machine drift of 1.1e-3 on decision_receipt.margin_percent
+# (5.2995 vs golden 5.3006). Candidate scores, already rounded to 4
+# decimals, drifted by 1 ulp (1.0e-4). Tolerance is the wider of
+# abs=2e-3 and rel=5e-4 so 1.1e-3 still passes while a 1e-2 mutation
+# (the reverse test below) fails.
+GOLDEN_FLOAT_ABS = 2e-3
+GOLDEN_FLOAT_REL = 5e-4
 
 
 def public_score_request() -> dict[str, Any]:
@@ -117,11 +139,102 @@ def _count_calls(monkeypatch, owner: Any, name: str, *, from_engine: bool = Fals
     return counter
 
 
+def _is_json_number(value: Any) -> bool:
+    return isinstance(value, (int, float)) and not isinstance(value, bool)
+
+
+def _assert_tiered_equal(actual: Any, expected: Any, *, path: str) -> None:
+    if isinstance(expected, dict):
+        assert isinstance(actual, dict), path
+        assert set(actual) == set(expected), f"{path} keys {set(actual)!r} != {set(expected)!r}"
+        for key in expected:
+            _assert_tiered_equal(actual[key], expected[key], path=f"{path}.{key}")
+        return
+    if isinstance(expected, list):
+        assert isinstance(actual, list), path
+        assert len(actual) == len(expected), f"{path} length {len(actual)} != {len(expected)}"
+        for index, (left, right) in enumerate(zip(actual, expected)):
+            _assert_tiered_equal(left, right, path=f"{path}[{index}]")
+        return
+    if isinstance(expected, int) and isinstance(actual, int) and not isinstance(expected, bool) and not isinstance(actual, bool):
+        assert actual == expected, f"{path}: {actual!r} != {expected!r}"
+        return
+    if _is_json_number(expected) or _is_json_number(actual):
+        left = float(actual)
+        right = float(expected)
+        tolerance = max(GOLDEN_FLOAT_ABS, GOLDEN_FLOAT_REL * abs(right))
+        assert math.isclose(left, right, rel_tol=0.0, abs_tol=tolerance), (
+            f"{path}: {left!r} vs {right!r} exceeds {tolerance}"
+        )
+        return
+    assert actual == expected, f"{path}: {actual!r} != {expected!r}"
+
+
+def _assert_memoization_payloads(actual: dict[str, Any], expected: dict[str, Any]) -> None:
+    _assert_tiered_equal(actual["candidate_scores"], expected["candidate_scores"], path="candidate_scores")
+    _assert_tiered_equal(actual["decision_receipt"], expected["decision_receipt"], path="decision_receipt")
+
+
+def _event_engine_request() -> dict[str, Any]:
+    """Dated events for compute_event_candidate_rows (not the v5 score payload)."""
+    return {
+        "birth_date": "1990-01-01",
+        "start_time": "12:00",
+        "end_time": "12:02",
+        "lat": 39.9,
+        "lon": 116.4,
+        "tz": 8,
+        "events": [
+            {
+                "id": "00000000-0000-4000-8000-000000000001",
+                "domain": "education",
+                "event_kind": "education_start",
+                "date": "2008-07-01",
+                "precision": "day",
+                "summary": "入学",
+            },
+            {
+                "id": "00000000-0000-4000-8000-000000000002",
+                "domain": "career",
+                "event_kind": "career_entry",
+                "date": "2012-06-15",
+                "precision": "day",
+                "summary": "入职",
+            },
+        ],
+    }
+
+
+def _contexts_with_layer_cache_cleared(contexts: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    cleared: list[dict[str, Any]] = []
+    for context in contexts:
+        clone = dict(context)
+        for key in CACHE_LAYER_KEYS:
+            clone[key] = None
+        cleared.append(clone)
+    return cleared
+
+
 def test_score_candidates_matches_baseline_golden() -> None:
     expected = json.loads(GOLDEN_PATH.read_text(encoding="utf-8"))
     actual = json.loads(json.dumps(_golden_payload(), ensure_ascii=True))
-    assert actual["candidate_scores"] == expected["candidate_scores"]
+    _assert_memoization_payloads(actual, expected)
-    assert actual["decision_receipt"] == expected["decision_receipt"]
+
+
+def test_golden_float_shift_of_1e_minus_2_fails() -> None:
+    expected = json.loads(GOLDEN_PATH.read_text(encoding="utf-8"))
+    mutated = json.loads(json.dumps(expected))
+    mutated["candidate_scores"][0]["score"] += 1e-2
+    with pytest.raises(AssertionError):
+        _assert_memoization_payloads(expected, mutated)
+
+
+def test_golden_discrete_field_mismatch_fails() -> None:
+    expected = json.loads(GOLDEN_PATH.read_text(encoding="utf-8"))
+    mutated = json.loads(json.dumps(expected))
+    mutated["candidate_scores"][0]["time"] = "99:99"
+    with pytest.raises(AssertionError):
+        _assert_memoization_payloads(expected, mutated)
 
 
 def test_shadbala_verified_fields_match_with_and_without_birth_minute() -> None:
@@ -372,3 +485,51 @@ def test_compute_event_candidate_rows_reuses_static_context_without_mutating_it(
     assert len(rows) == 2
     assert [frozenset(context) for context in contexts] == original_keys
     assert all("_transit_chart_cache" not in context for context in contexts)
+
+
+def test_cached_static_context_matches_uncached_fallback(monkeypatch) -> None:
+    request = _event_engine_request()
+    cached = [
+        build_candidate_static_context(request, candidate)
+        for candidate in _candidate_datetimes(request)
+    ]
+    assert len(cached) >= 2
+    assert len(request["events"]) >= 2
+    for context in cached:
+        for key in CACHE_LAYER_KEYS:
+            assert context.get(key) is not None
+    uncached = _contexts_with_layer_cache_cleared(cached)
+    for context in uncached:
+        for key in CACHE_LAYER_KEYS:
+            assert context[key] is None
+    calls = _count_calls(monkeypatch, event_engine.shadbala, "calc_shadbala")
+    rows_cached = compute_event_candidate_rows(request, static_contexts=cached)
+    cached_calls = calls[0]
+    rows_uncached = compute_event_candidate_rows(request, static_contexts=uncached)
+    uncached_calls = calls[0] - cached_calls
+    assert rows_cached == rows_uncached
+    assert cached_calls == 0
+    assert uncached_calls > cached_calls
+    assert uncached_calls >= len(cached) * len(request["events"])
+
+
+def test_poisoned_static_cache_diverges_from_live_rows() -> None:
+    request = _event_engine_request()
+    cached = [
+        build_candidate_static_context(request, candidate)
+        for candidate in _candidate_datetimes(request)
+    ]
+    rows_cached = compute_event_candidate_rows(request, static_contexts=cached)
+    poisoned = [dict(context) for context in cached]
+    poisoned[0]["shadbala_result"] = {
+        "planets": {
+            name: {
+                "sthana_bala": {"total": 0.0},
+                "drik_bala": 0.0,
+                "naisargika_bala": 0.0,
+            }
+            for name in ("Sun", "Moon", "Mars", "Mercury", "Jupiter", "Venus", "Saturn")
+        }
+    }
+    rows_poisoned = compute_event_candidate_rows(request, static_contexts=poisoned)
+    assert rows_poisoned != rows_cached