From 35de346f3f7a8f38a5a0b48eedb9c0db964a8782 Mon Sep 17 00:00:00 2001
From: shaofl <2899218482@qq.com>
Date: Tue, 28 Jul 2026 02:08:33 -0700
Subject: [PATCH 1/3] test(eval): bound the bounded warmup class by its
 structural cell ceiling (D5 C4a review LOW-1)

---
 qt/factor_eval_reconcile.py         | 52 +++++++++++++++-
 tests/test_factor_eval_reconcile.py | 95 +++++++++++++++++++++++++++++
 2 files changed, 145 insertions(+), 2 deletions(-)

diff --git a/qt/factor_eval_reconcile.py b/qt/factor_eval_reconcile.py
index 4ca0352..0ea507d 100644
--- a/qt/factor_eval_reconcile.py
+++ b/qt/factor_eval_reconcile.py
@@ -23,7 +23,18 @@
      * bounded factor: the differing row sits in the grid's first
        ``lookback_depth - 1`` trading dates — ALL THREE directions allowed
        (frozen-NaN -> new-finite, finite -> finite as a partial pool fills,
-       and a finite value on a row the frozen panel does not have);
+       and a finite value on a row the frozen panel does not have). The
+       class is CEILED at ``(lookback_depth - 1) x |frozen symbols|`` cells
+       (D5 C4a review LOW-1): a structural bound, not a calibrated one —
+       every warmup cell is a distinct (date, symbol) pair with the date in
+       the warmup boundary and the symbol known to the frozen panel (the
+       ``new_only_finite`` direction requires ``frozen_symbols`` membership),
+       so the class cannot exceed it by construction and an overshoot means
+       the boundary/counting logic itself is broken -> FAIL. AMPLITUDE stays
+       unbounded inside the region by the standing ruling: for C5's purpose
+       (geometry reconciliation) that is acceptable, because the two loading
+       geometries legitimately differ on exactly those cells — the value gate
+       for the rest of the grid is the 1e-12 tolerance outside the boundary;
      * valid-day-POOLED factor: the early region [2021-07-01, 2021-10-31],
        same three directions, and the per-month counts must be
        non-increasing ("按月递减至零"; a violation fails the mode) with ONE
@@ -731,6 +742,11 @@ class PanelDiff:
     warmup_by_direction: dict[str, int] = field(default_factory=dict)
     warmup_monotonic: bool = True
     warmup_exempt_month: str | None = None
+    #: Structural cell ceiling of the bounded ``warmup_left_extension``
+    #: class — ``(lookback_depth - 1) x |frozen symbols|`` (None for pooled
+    #: factors, whose class is gated by the monthly monotonicity machinery
+    #: instead). See ``classify_panel_differences``.
+    warmup_max_cells: int | None = None
     ok: bool = True
 
     def by_class(self, classification: str) -> list[PanelCellDiff]:
@@ -771,6 +787,20 @@ def classify_panel_differences(
     where diffs land; a frozen panel with no finite value gets no
     exemption (conservative). Later months stay gated.
 
+    Bounded warmup class CELL CEILING (D5 C4a review LOW-1): the class is
+    capped at ``(lookback_depth - 1) x |frozen symbols|`` cells. This is a
+    STRUCTURAL bound, not a calibrated one: every cell the classifier can
+    place in the class is a distinct (date, symbol) pair whose date sits in
+    the warmup boundary and whose symbol is known to the frozen panel (the
+    ``new_only_finite`` direction explicitly requires frozen-symbol
+    membership), so the class cannot exceed the ceiling by construction —
+    an overshoot means the boundary/counting logic itself is broken and
+    fails the mode. AMPLITUDE inside the region stays UNBOUNDED by the
+    standing ruling: for C5's geometry-reconciliation purpose that is
+    acceptable, because the two loading geometries legitimately differ on
+    exactly those cells (the old runner's partial pool vs the saturated
+    one); everything outside the boundary is still gated at 1e-12.
+
     ``is_cross_sectional`` (D4c ``stores_intermediate`` factors — the served
     value is a per-date cross-sectional combine): enables the
     ``threshold_flip_contamination`` class and the wider float-tail cap
@@ -801,6 +831,18 @@ def classify_panel_differences(
     n_warm = max(int(lookback_depth) - 1, 0)
     warmup_dates = set(grid_dates[:n_warm])
 
+    # Structural cell ceiling of the bounded warmup class (D5 C4a review
+    # LOW-1): every warmup cell the classifier can emit is a distinct
+    # (date, symbol) pair with the date in ``warmup_dates`` and the symbol
+    # in the frozen panel's symbol set (the ``new_only_finite`` direction
+    # requires frozen-symbol membership), so the class cannot exceed
+    # ``n_warm x |frozen symbols|`` by construction — an overshoot means
+    # the boundary/counting logic is broken and fails the mode. Pooled
+    # factors get NO count ceiling: their class is gated by the monthly
+    # monotonicity + exempt-month machinery instead.
+    if not is_pooled:
+        result.warmup_max_cells = n_warm * len(frozen["symbol"].unique())
+
     def _in_warmup(date: pd.Timestamp, symbol: str) -> bool:
         if is_pooled:
             return early_lo <= date <= early_hi
@@ -944,6 +986,11 @@ def _warmup_cell(date, symbol, frozen_v, new_v, direction: str) -> None:
         not any(d.classification.startswith("unclassified") or
                 d.classification.startswith("unregistered") for d in result.diffs)
         and result.warmup_monotonic
+        and (
+            result.warmup_max_cells is None
+            or len(result.by_class("warmup_left_extension"))
+            <= result.warmup_max_cells
+        )
         and len(result.by_class("float_reordering_tail")) <= float_tail_max
         and len(result.by_class("threshold_flip_tail")) <= flip_max
         and len(result.by_class("threshold_flip_contamination"))
@@ -1120,12 +1167,13 @@ def run_panels_mode(config_path: str, factor_id: str, repo_root: Path) -> PanelD
     )
     logger.info(
         "panels %s: rows frozen=%d new=%d equal=%d tol=%d warmup=%d(%s) "
-        "exempt_month=%s "
+        "warmup_ceiling=%s exempt_month=%s "
         "float_tail=%d threshold_flip=%d flip_contamination=%d footprint=%d "
         "unclassified=%d max_rel=%.3e live_calls=%d ok=%s",
         factor_id, result.rows_frozen, result.rows_new, result.equal,
         result.within_tolerance, len(result.by_class("warmup_left_extension")),
         result.warmup_by_direction,
+        result.warmup_max_cells,
         result.warmup_exempt_month,
         len(result.by_class("float_reordering_tail")),
         len(result.by_class("threshold_flip_tail")),
diff --git a/tests/test_factor_eval_reconcile.py b/tests/test_factor_eval_reconcile.py
index c70c475..714edeb 100644
--- a/tests/test_factor_eval_reconcile.py
+++ b/tests/test_factor_eval_reconcile.py
@@ -665,6 +665,101 @@ def test_panels_bounded_new_only_finite_row_outside_warmup_fails():
     assert len(result.by_class("unregistered_new_finite_row")) == 1
 
 
+# --------------------------------------------------------------------------- #
+# Bounded warmup CELL CEILING (D5 C4a review LOW-1): the class is capped at
+# (lookback_depth - 1) x |frozen symbols| — a structural bound (every class
+# cell is a distinct (date, symbol) pair in warmup_dates x frozen_symbols),
+# so an overshoot means the boundary/counting logic itself is broken.
+# --------------------------------------------------------------------------- #
+def test_panels_bounded_warmup_ceiling_is_the_structural_derivation():
+    # 3 symbols, lookback_depth=4 -> ceiling = 3 warmup dates x 3 symbols = 9.
+    frozen = _frozen_panel(
+        [(f"2021-07-0{d}", s, 1.0) for d in (1, 2, 3, 4, 5) for s in ("A", "B", "C")],
+        "f",
+    )
+    new = _new_series(
+        [(f"2021-07-0{d}", s, 1.0) for d in (1, 2, 3, 4, 5) for s in ("A", "B", "C")]
+    )
+    result = classify_panel_differences(
+        new, frozen, factor_id="f", is_pooled=False, lookback_depth=4
+    )
+    assert result.ok, result.diffs
+    assert result.warmup_max_cells == 3 * 3
+
+
+def test_panels_bounded_warmup_cells_exactly_at_the_ceiling_pass():
+    # lookback_depth=3 -> 2 warmup dates; 2 symbols -> ceiling 4. All four
+    # warmup cells move (finite -> finite): AT the ceiling, still registered.
+    frozen = _frozen_panel(
+        [
+            ("2021-07-01", "A", 0.10), ("2021-07-01", "B", 0.50),
+            ("2021-07-02", "A", 0.20), ("2021-07-02", "B", 0.60),
+            ("2021-07-03", "A", 0.30), ("2021-07-03", "B", 0.70),
+        ],
+        "f",
+    )
+    new = _new_series(
+        [
+            ("2021-07-01", "A", 0.11), ("2021-07-01", "B", 0.55),
+            ("2021-07-02", "A", 0.25), ("2021-07-02", "B", 0.65),
+            ("2021-07-03", "A", 0.30), ("2021-07-03", "B", 0.70),
+        ]
+    )
+    result = classify_panel_differences(
+        new, frozen, factor_id="f", is_pooled=False, lookback_depth=3
+    )
+    assert result.ok, result.diffs
+    assert len(result.by_class("warmup_left_extension")) == 4
+    assert result.warmup_max_cells == 4
+
+
+def test_panels_bounded_warmup_cells_beyond_the_ceiling_fail():
+    # REVERSE (LOW-1): a frozen panel with DUPLICATE (date, symbol) rows
+    # emits one warmup cell per duplicate — 4 cells against the structural
+    # ceiling of 2 warmup dates x 1 symbol = 2. An overshoot of a bound the
+    # class cannot structurally exceed means the counting is broken -> FAIL.
+    # The individual cells still classify as warmup (the ceiling is a COUNT
+    # gate on the class, not a per-cell reclassification).
+    frozen = _frozen_panel(
+        [
+            ("2021-07-01", "A", 0.10), ("2021-07-01", "A", 0.10),  # duplicate
+            ("2021-07-02", "A", 0.20), ("2021-07-02", "A", 0.20),  # duplicate
+            ("2021-07-03", "A", 0.30),
+        ],
+        "f",
+    )
+    new = _new_series(
+        [("2021-07-01", "A", 0.15), ("2021-07-02", "A", 0.25), ("2021-07-03", "A", 0.30)]
+    )
+    result = classify_panel_differences(
+        new, frozen, factor_id="f", is_pooled=False, lookback_depth=3
+    )
+    assert len(result.by_class("warmup_left_extension")) == 4
+    assert result.warmup_max_cells == 2
+    assert not result.ok
+
+
+def test_panels_pooled_warmup_class_has_no_cell_ceiling():
+    # Pooled factors: NO count ceiling — the monthly monotonicity gate is
+    # the mechanism (a large but non-increasing count still passes).
+    frozen = _frozen_panel(
+        [
+            ("2021-07-01", "A", 1.0), ("2021-07-01", "B", 1.0),
+            ("2021-08-02", "A", 1.0),
+        ],
+        "f",
+    )
+    new = _new_series(
+        [("2021-07-01", "A", 2.0), ("2021-07-01", "B", 2.0), ("2021-08-02", "A", 2.0)]
+    )
+    result = classify_panel_differences(
+        new, frozen, factor_id="f", is_pooled=True, lookback_depth=20
+    )
+    assert result.ok, result.diffs
+    assert len(result.by_class("warmup_left_extension")) == 3
+    assert result.warmup_max_cells is None
+
+
 def test_panels_pooled_warmup_class_all_directions_with_monthly_decrease():
     frozen = _frozen_panel(
         [

From daf173b1335fed668dbe18040911d4702830236d Mon Sep 17 00:00:00 2001
From: shaofl <2899218482@qq.com>
Date: Tue, 28 Jul 2026 02:08:33 -0700
Subject: [PATCH 2/3] docs(factors): D5 C5 property-test migration map + audit
 report skeleton

---
 docs/factors/d5_c5_audit_report.md            | 131 ++++++++++++++
 .../factors/d5_property_test_migration_map.md | 170 ++++++++++++++++++
 2 files changed, 301 insertions(+)
 create mode 100644 docs/factors/d5_c5_audit_report.md
 create mode 100644 docs/factors/d5_property_test_migration_map.md

diff --git a/docs/factors/d5_c5_audit_report.md b/docs/factors/d5_c5_audit_report.md
new file mode 100644
index 0000000..f373443
--- /dev/null
+++ b/docs/factors/d5_c5_audit_report.md
@@ -0,0 +1,131 @@
+# D5 C5 —— 四腿全量对账审计报告
+
+> **状态**：**骨架**。结果区全部 TBD，由全量对账（11 因子 × panels/reports/anchors）
+> 跑完后填充。填充纪律：每处差异必须落在 §二 预登记清单的具名条目内；**未编目的
+> 差异 = 验收不通过**（设计 v3.2 §五腿 3：「允许差异，但每处差异必须归因到具名原因；
+> 变好的差异优先怀疑泄漏」）。
+> **执行环境**：11 因子 × {no_book, with_book} 统一 exec-only runner
+> （`qt/factor_eval_runner.py`）+ 对账 harness（`qt/factor_eval_reconcile.py` 三模式）。
+
+---
+
+## 一、范围与方法
+
+### 1.1 四腿定义（设计 v3.2 §五）
+
+| 腿 | 内容 | 本步载体 |
+|---|---|---|
+| 腿 1 性质测试搬迁 | 旧 runner 测试的每条性质在新套件有映射，覆盖数非降（R16） | `docs/factors/d5_property_test_migration_map.md`（59/59 有映射，0 静默丢失） |
+| 腿 2 手算锚 | 每因子 ≥5 行从 raw parquet 徒手复算，不 import 新引擎；warmup 区差异归 `warmup_left_extension`，jump 非 warmup 行必须 reconcile（证明 service 携带截断定义） | harness `--mode anchors` vs `artifacts/refactor_baseline/hand_anchors_d2.json` |
+| 腿 3 语义对账 | 新 artifact vs 冻结 exec artifact **值级**比对（IC/ICIR/分位价差/verdict；逐字节已知不成立，见 §二 #1）；差异按白名单分类，聚合路径之外任何变化即失败 | harness `--mode reports`（JSON 叶子 + MD 行级配对，strict/no-book 与 bookclose 双闸） |
+| 腿 4 面板腿 | service 面板（decision 视图 × exec_to_exec）vs D1 冻结面板**逐格**（1e-12；NaN 集合差异单独分类）；jump 走 `pr_c_cutoff_fix` 参照物 | harness `--mode panels`（具名类：`warmup_left_extension` / `float_reordering_tail` / `threshold_flip_tail` / `threshold_flip_contamination`） |
+
+### 1.2 硬闸门与基线纪律
+
+- **77/77 硬闸门**：三个模式入口都先跑 `verify_all()`，冻结 exec 基线字节不可达或被改
+  是**硬错误不是 skip**（没有基线就没有对账；「对账通过但没有基线」是
+  compare_postmerge 空对账形态）。
+- **只读冻结基线**：对账只读 D5a 冻结的 77 份（`artifacts/refactor_baseline/exec_baseline/`，
+  sha256 manifest 入 git），**绝不读 live 路径**——新 runner 写同名文件，跑一次就毁掉
+  唯一基线。
+- **cache-only**：每次 service 取数后断言 `stk_mins_live_calls == 0`，非零即 ABORT。
+- **harness 加固（本准备步）**：bounded `warmup_left_extension` 类加**结构性 cell 数
+  ceiling** =（w−1 个 warmup 交易日）×（冻结面板 symbol 数）——该类按定义不可能超过
+  此界，越界即 FAIL（C4a 评审 LOW-1；类内**幅度仍不设限**的既有裁定不变，理由：两种
+  加载几何在该区本来就合法地不同，区外仍由 1e-12 把门）。
+
+## 二、预登记漂移清单（对账白名单）
+
+以下来源：编目 §七 / §七之二 / §七之三 / §七之四 / §七之五
+（`docs/factors/d5_runner_difference_catalogue.md`）+ 交接 §3
+（`tmp/design/HANDOFF_2026-07-26_refactor_execution.md`）。**不在本表内的差异一律
+unclassified → FAIL。**
+
+| # | 名称 | 边界 | 出处 |
+|---|---|---|---|
+| 1 | 契约 v1.0/v1.1 artifact 新增（`spec` 16→20 键 / `eval_config` +3 键 / 顶层 `eval_contract_version` + `corrections` / MD +4 行 provenance + 每条更正 +1 行） | 全为**新增**、零删除、零数值变化；逐字节对账已知不成立，只成立值级 + 结构性断言 | 编目 §七、§七之二；交接 §3 #1 |
+| 2 | exec 侧 no_book 与 with_book 的 `eval_config` 首次合法不同（`book_view`: null vs run 模式） | 仅此一键；不得当作两格之间的回归 | 编目 §七之二 ⚠️；交接 §3 #2 |
+| 3 | served-panel NaN 足迹（D4c）：行集变密、按行数算分母的覆盖率变，读值不受影响 | 额外行必须全 NaN；有限值的新行只许落在 warmup 边界内 | 交接 §3 #3 |
+| 4 | `warmup_left_extension`（panels/anchors 通用，三方向） | bounded：冻结网格前 w−1 个交易日（w=`lookback_depth`），cell 数 ≤ (w−1)×|symbols|（本步新增 ceiling）；pooled：早区 [2021-07-01, 2021-10-31] + 按月计数非递增，豁免月 = 冻结面板**首个 finite 值所在月**（结构性锚定；全 NaN 冻结面板不豁免） | 编目 §七之三根因一、§七之四 #2、§七之五裁定 1；交接 §3 #4② |
+| 5 | `float_reordering_tail` | rel ≤ 5e-12 或 abs ≤ 1e-12；cap：bars-only ≤101 cell，截面 ≤1,000 cell | 编目 §七之三根因二、§七之五裁定 3 |
+| 6 | `threshold_flip_tail`（count 因子） | 幅度恰 ±1 count、rel ≤ 1e-2、≤25 cell；实测 600623.SH 2023-06-15..07-14 | 编目 §七之三根因二 |
+| 7 | `threshold_flip_contamination`（仅截面因子） | 窗口 [2023-06-01, 2023-07-14]；直接 symbol 600623.SH ≤2e-04，其余 ≤1e-06，总 ≤20,000 cell；逐格校验，越界即 unclassified | 编目 §七之五裁定 2 |
+| 8 | `warmup_aggregate_effect`（reports 腿） | 仅聚合叶子（`sections[*]` / `verdict.reasons[*]` / `verdict.axes.*`）且**数字承载**；纯标签翻转（PASS→FAIL 无数字）仍失败 | 编目 §七之三 reports #2 |
+| 9 | `registered_sanity_stem_rename` | sanity 报告名 `eval_<name>_` → `factor_eval_<factor_id>_`，双侧以 `_exec_basis_sanity.md` 收尾 | 编目 §七之三 reports #4 |
+| 10 | `registered_run_order_artifact` | `exec_price_artifact_reused` False→True（同会话 run 序）；反向不登记 | 编目 §七之三 reports #5 |
+| 11 | jump 更正效应（`registered_correction_effect`） | 仅当新 JSON 带 `corrections` 结构化承载（契约 v1.1）；其值级验证在 panels 腿（`pr_c_cutoff_fix` 参照物）与 anchors 腿（非 warmup 行必须 reconcile），不在 reports 腿 | 编目 §七之三 reports #6、§七之二 ✅；交接 §3 #4③ |
+| 12 | vpq `neutralization_coverage` add-Section | 整个 subtree 为新增；按 section **名字**匹配（`REGISTERED_EXTRA_SECTIONS`），绝不按裸下标 | 编目 §七之四 #1 |
+| 13 | `book_view_effect`（with_book(decision) 格） | decision 书 vs 冻结 close 书的 Incremental 轴数值差：**全量报告、不设闸**；闸在 `_bookclose` 格（见 §四） | harness 设计（a)/(b) 分解；交接 §3 D5 C4 ⚠️ |
+| 14 | `hand_anchors_d2.json` 报 `all_ok_frozen14: False` | 70 行 5 处失配**全部且仅有 jump**（手算已截断 vs 冻结 `panels_d2` 未截断）——正确信号非回归 | 交接 §3 #5 |
+
+## 三、每因子结果表（TBD）
+
+填写口径：panels = 具名类 cell 计数 + unclassified 计数（必须为 0）；reports =
+no_book / with_book(decision) / with_book(bookclose) 三格各自的登记类计数与
+unregistered 计数（strict 格必须为 0）；anchors = ok/warmup/failed 行数（failed 必须
+为 0）。`stk_mins_live_calls` 每因子必须为 0。
+
+| 因子 | panels: warmup / float_tail / flip_tail / flip_contam / footprint / unclassified / ceiling | reports: no_book | reports: with_book(decision) | reports: with_book(bookclose) | anchors: ok / warmup / failed | live_calls |
+|---|---|---|---|---|---|---|
+| jump_amount_corr_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| minute_ideal_amp_10 | TBD | TBD | TBD | TBD | TBD | TBD |
+| amp_marginal_anomaly_vol_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| volume_peak_count_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| intraday_amp_cut_10 | TBD | TBD | TBD | TBD | TBD | TBD |
+| peak_interval_kurtosis_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| valley_relative_vwap_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| valley_ridge_vwap_ratio_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| ridge_minute_return_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| valley_price_quantile_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+| peak_ridge_amount_ratio_20 | TBD | TBD | TBD | TBD | TBD | TBD |
+
+## 四、with_book 差异分解表（TBD）
+
+交接 §3 的硬要求：with_book 的差异必须可分解、**整体归因不可接受**（「with_book 变了
+是因为书视图改了」会把引擎回归藏在 intended change 后面）。三分量：
+
+- **A = no_book 引擎效应**：no_book 格的登记类差异（干净的「有没有弄坏东西」测试，
+  先对干净）。
+- **B = close 书引擎效应**：`with_book(bookclose)` 格（legacy-faithful close 书，strict）
+  的登记类差异——引擎在带书条件下的效应，与 A 同口径可比。
+- **C−B = 书视图修正**：`with_book(decision)` 格的 `book_view_effect` 叶子（decision 书
+  vs close 书的预期差异，全量列出不设闸）——即 §1.1 活缺陷（close(d) 书）的修正幅度。
+
+| 因子 | A（no_book 各类计数） | B（bookclose 各类计数） | C−B（book_view_effect 叶子数 / 涉及轴） | 分解闭合？（B 超 A 的部分是否全部具名） |
+|---|---|---|---|---|
+| jump_amount_corr_20 | TBD | TBD | TBD | TBD |
+| minute_ideal_amp_10 | TBD | TBD | TBD | TBD |
+| amp_marginal_anomaly_vol_20 | TBD | TBD | TBD | TBD |
+| volume_peak_count_20 | TBD | TBD | TBD | TBD |
+| intraday_amp_cut_10 | TBD | TBD | TBD | TBD |
+| peak_interval_kurtosis_20 | TBD | TBD | TBD | TBD |
+| valley_relative_vwap_20 | TBD | TBD | TBD | TBD |
+| valley_ridge_vwap_ratio_20 | TBD | TBD | TBD | TBD |
+| ridge_minute_return_20 | TBD | TBD | TBD | TBD |
+| valley_price_quantile_20 | TBD | TBD | TBD | TBD |
+| peak_ridge_amount_ratio_20 | TBD | TBD | TBD | TBD |
+
+## 五、判定（TBD）
+
+| 因子 | 腿 1 性质映射 | 腿 2 anchors | 腿 3 reports | 腿 4 panels | 判定 |
+|---|---|---|---|---|---|
+| （11 行，每行 PASS/FAIL + 一句话归因） | 59/59（本准备步已立） | TBD | TBD | TBD | TBD |
+
+**整体结论**：TBD（全 11 因子四腿全过才进入 C6 删除旧 runner；任一不过则旧 runner
+不删、迭代新 runner——设计 v3.2 §九 D5 行回滚条款）。
+
+## 六、已知局限（登记，不阻塞）
+
+1. **C4a 评审 LOW-1（已处置）**：bounded `warmup_left_extension` 类原无 cell 数上限，
+   warmup 区内任意幅度错误都被吞（暴露面 ~1.6% 网格）。本准备步已加结构性 ceiling
+   （§1.2）；**类内幅度仍不设限**是既有裁定——C5 用途（几何对账）下可接受，因为两种
+   加载几何在该区本来就合法地不同；区外由 1e-12 把门。
+2. **C4a 评审 LOW-2**：reports 腿 `warmup_aggregate_effect` 的「数字承载」是 digit OR
+   规则——散文理由行的语义可变（改了词但含数字仍放行）；verdict 标签等无数字翻转
+   仍严格失败。判定阅读时应以 panels 腿为值级闸门（reports 腿是聚合层）。
+3. **NIT 登记**：C4a NIT-1/NIT-2（装饰性，不影响判定）。
+4. **既有披露缺陷（非本次回归，见交接 §4）**：六个 JSON 字段被 `MAX_VALUE_CHARS=200`
+   静默截断（census 守卫登记为 frozen set）；9/11 dashboard FACTOR DEFINITION 带叠字
+   （限行 + 省略标记已落地，C5 重渲染后残留形态需目视确认）。
+5. **census 守卫 cwd 陷阱**：读真实语料的 census 测试在无语料的 worktree 里静默
+   skip——全量对账的 gate 行必须报「语料可达/不可达」两组数字，别只看总数。
diff --git a/docs/factors/d5_property_test_migration_map.md b/docs/factors/d5_property_test_migration_map.md
new file mode 100644
index 0000000..0a07f03
--- /dev/null
+++ b/docs/factors/d5_property_test_migration_map.md
@@ -0,0 +1,170 @@
+# D5 C5 腿 1 —— 性质测试搬迁映射表（旧 runner 测试 → 新套件）
+
+> **状态**：D5 C5 准备交付物（设计 v3.2 §五腿 1「性质测试搬迁」的映射底账；C6 的 R16
+> 「覆盖数非降」审计直接消费本表）。
+> **口径**：旧套件 = 11 个旧 runner 的测试文件 `tests/test_eval_*_runner.py`（**实有 10 份、
+> 59 条测试函数**——`peak_ridge_amount_ratio`（PR-M）**没有** runner 测试文件，编目
+> §六 BUG 3 已登记，见 §五）。新套件 = 统一 runner 落地后的测试面（`test_factor_eval_runner.py`
+> / `test_factor_eval_disclosures.py` / `test_factor_eval_providers.py` /
+> `test_factor_eval_reconcile.py` / `test_exec_basis_eval.py` / `test_factor_service.py` /
+> `test_factor_store_universe.py` / `test_minute_binding_vpq.py` /
+> `test_minute_diagnostics_channel.py` / 各 `test_<factor>_factor.py` /
+> `test_factor_eval_contract.py` / `test_factor_eval_standard.py` /
+> `test_factor_requires_spec_v1.py` 等）。
+> **基线**：`feat/factor-refactor-d5-c5-prep` @ origin/main（含 PR #105/#107）。全部 node-ID
+> 在该基线上以 `pytest --collect-only -q` 核对存在（抽查记录见 §六）。
+> **枚举方式**：`pytest --collect-only -q tests/test_eval_*_runner.py` 得每文件计数，再逐条
+> `grep -n "def test"` 人工归类。计数：4+4+4+4+4+10+14+4+7+4 = **59**。
+
+## 状态定义
+
+| 状态 | 含义 |
+|---|---|
+| **已搬迁** | 测试体逐字（或近逐字）移入新文件；node-ID 改变，断言不变 |
+| **已覆盖** | 旧测试主张的**性质**在新套件由一条或多条测试承担（载体可不同） |
+| **载体退役** | 旧测试的**字面断言形态**不再存在，但其性质在新层有钉；理由逐条写明 |
+| **NET-NEW** | 新套件超出旧套件的性质（旧套件从未测过） |
+
+---
+
+## 一、六个同构 4 测试文件（24 条）
+
+适用文件（每文件 4 条，结构逐字同构，下表以 `test_eval_jump_amount_corr_runner.py`
+的 node-ID 为代表，其余五个文件的同名测试映射相同）：
+
+- `tests/test_eval_jump_amount_corr_runner.py`（PR-C，4）
+- `tests/test_eval_minute_ideal_amplitude_runner.py`（PR-D，4）
+- `tests/test_eval_amp_marginal_anomaly_vol_runner.py`（PR-E，4）
+- `tests/test_eval_volume_peak_count_runner.py`（PR-F，4）
+- `tests/test_eval_peak_interval_kurtosis_runner.py`（PR-H，4）
+- `tests/test_eval_valley_relative_vwap_runner.py`（PR-I，4）
+
+| 旧 node-ID（jump 代表） | 性质 | 新套件映射 | 状态 |
+|---|---|---|---|
+| `test_minute_loader_is_cache_only_and_discloses_empty` | 分钟取数 cache-only（`live_calls=0`）；未覆盖 symbol 披露为 empty、绝不暖跑；工程化 fixture 上的手算值逐格对上 | `tests/test_factor_eval_providers.py::test_minute_bars_reads_normalized_bars_with_zero_live_calls`、`::test_minute_bars_empty_and_missing_are_empty_never_live`（cache-only + 空披露）；`tests/test_factor_store_universe.py::test_second_symbol_batch_is_filled_not_silently_dropped[fid]`（缺口按 (date,symbol) 判定、不静默丢名）；手算值 → 各因子自己的 `tests/test_<factor>_factor.py`（D2 迁移，11 因子一文件，如 `test_jump_amount_corr_factor.py` 16 条含手算锚） | 已覆盖 |
+| `test_minute_loader_blocks_when_nothing_cached` | 全空缓存 → loud `ValueError("no requested symbol produced")`，绝不静默出空结果 | `tests/test_factor_service.py::test_zero_output_symbol_is_not_silently_dropped`、`::test_zero_output_symbol_veto_mutation`（零输出否决 + mutation，**更强**：逐 symbol 级别）；`tests/test_factor_eval_providers.py::test_minute_bars_empty_and_missing_are_empty_never_live` | 已覆盖 |
+| `test_evaluate_two_runs_produces_full_reports_and_incremental_axis` | 双跑（no_book/with_book）产出 8 个 mandatory sections + verdict；no_book purity=Skipped → Incremental 轴 NOT_ASSESSED；with_book purity=Section 含 `incremental_ic_ir` → 轴被评估；md/json/dashboard 全落盘；metrics 提取出门禁字段 | `tests/test_exec_basis_eval.py::test_exec_basis_eval_writes_its_own_reports_and_leaves_the_control_alone`、`::test_the_two_exec_artifacts_state_their_own_information_sets`、`::test_exec_basis_eval_discloses_parameters_and_coverage_in_every_report`（落盘 + 披露）；`tests/test_factor_eval_standard.py::test_all_eight_sections_are_present_and_ordered`（8 节齐备）；Incremental 轴语义 → `tests/test_factor_eval_contract.py` 的 purity/Incremental 判决测试（契约层 v0.7 起未动）；metrics → `qt/exec_basis_eval.py::_extract_metrics` 由 `test_exec_basis_eval.py` 各测试间接钉住（metrics dict 喂 CLI 行，见 `test_exec_basis_cli_line_survives_an_absent_metric`） | 已覆盖 |
+| `test_no_book_run_never_reaches_adopt` | 结构性保证：no-book + 无执行事实 → 封顶 Watch（exploratory cap + NOT_ASSESSED 轴） | `tests/test_factor_eval_contract.py::test_exploratory_run_is_capped_at_watch_instead_of_adopt`（exploratory 封顶）；同文件的 Adopt 路径测试（`..._ADOPT_FACTS` 族）钉住 Tradable NOT_ASSESSED 时不可达 Adopt。**契约层在重构中一字未动**，此性质不需要 runner 级重测 | 已覆盖 |
+
+## 二、`tests/test_eval_ridge_minute_return_runner.py`（PR-K，10 条）
+
+| 旧 node-ID | 性质 | 新套件映射 | 状态 |
+|---|---|---|---|
+| `test_minute_loader_is_cache_only_and_discloses_empty` | 同 §一 #1 + CASE A 手算 ridge-minute return 日和（+2.5，区分求和 vs 复利） | §一 #1 各条 + `tests/test_ridge_minute_return_factor.py`（手算值测试族） | 已覆盖 |
+| `test_minute_loader_reports_the_ridge_scarcity_distribution` | ridge 分布 + 有效率必须**实测**于真实 bar；披露的 floor 必须是本 run 实际施加的门（否则披露在描述没人执行的门）；render 行含门值 | 诊断帧产生 → `tests/test_minute_diagnostics_channel.py::test_the_diagnostic_set_is_exactly_the_declared_three`、`::test_the_frames_carry_the_symbol_label_the_summarizers_read[factor_id]`、`::test_the_sink_is_scoped_to_the_emit_window[factor_id]`；汇总+门值 → `tests/test_factor_eval_disclosures.py::test_ridge_return_summarize_*`（3 条）+ `::test_summarizer_defaults_are_the_gates_the_binding_applies`（汇总器默认 ≡ 计算门 ≡ 模块常量，防手抄漂移） | 已覆盖 |
+| `test_coverage_reports_the_return_guard_attrition` | ridge bar 被日内 lag 丢弃必须**可见**（return_guard_attrition），不得静默吸收 | `tests/test_factor_eval_disclosures.py::test_ridge_return_summarize_reports_the_return_guard_attrition` | 已搬迁 |
+| `test_summarize_ridge_coverage_counterfactual_at_the_comparison_floor` | 反事实：把 ridge floor 提到比较基准（20）还剩几天 → 降门买到的天数是个数不是口号 | `tests/test_factor_eval_disclosures.py::test_ridge_return_summarize_counterfactual_at_the_comparison_floor` | 已搬迁 |
+| `test_summarize_ridge_coverage_handles_no_frames` | 空输入不除零、NaN 渲染 | `tests/test_factor_eval_disclosures.py::test_ridge_return_summarize_handles_no_frames`（**改名**：编目 BUG 6 的重名冲突在此家闭合，加因子前缀） | 已搬迁 |
+| `test_minute_loader_blocks_when_nothing_cached` | 同 §一 #2 | 同 §一 #2 | 已覆盖 |
+| `test_evaluate_two_runs_produces_full_reports_and_incremental_axis` | 同 §一 #3 | 同 §一 #3 | 已覆盖 |
+| `test_extract_metrics_surfaces_the_pr_k_comparison_quantities` | 头对头比较量（turnover / net_long_short_by_cost {1,2,4} / rank_autocorr / half_life / 截面规模 / monotonicity / gross spread）以数字形式可取 | **性质层**：这些量由评估核产出并钉在契约/标准层（`tests/test_factor_eval_standard.py::test_cost_scenarios_only_move_the_cost_line`、`::test_turnover_charges_the_first_period_and_a_full_rotation` 等；报告 JSON payload 含全量）。**载体**：旧 runner 私有 `extract_metrics` 的**键集形状**不搬迁——统一 runner 的 `_extract_metrics`（`qt/exec_basis_eval.py:242`）按编目 §四收敛为单一公共子集，比较量留在报告 JSON/section payload 里读 | 载体退役（理由 R1，见 §四） |
+| `test_absent_stability_metrics_render_as_na_not_a_crash` | Skipped section → 指标 None → 汇总行渲染 `n/a` 而非在错误处理之外抛 TypeError | `tests/test_exec_basis_eval.py::test_exec_basis_cli_line_survives_an_absent_metric`（同一性质，钉在新 CLI 行 `format_exec_basis_line` 上）。注：旧测试里对 `qt.cli._fmt_metric` 的直接断言随旧 CLI 命令在 C6 一并删除，性质已由新钉承担 | 已覆盖 |
+| `test_no_book_run_never_reaches_adopt` | 同 §一 #4 | 同 §一 #4 | 已覆盖 |
+
+## 三、`tests/test_eval_valley_ridge_vwap_ratio_runner.py`（PR-J，7 条）
+
+| 旧 node-ID | 性质 | 新套件映射 | 状态 |
+|---|---|---|---|
+| `test_minute_loader_is_cache_only_and_discloses_empty` | 同 §一 #1 + CASE A 手算 valley/ridge VWAP 比值 | §一 #1 各条 + `tests/test_valley_ridge_vwap_ratio_factor.py` 手算值族 | 已覆盖 |
+| `test_minute_loader_reports_the_ridge_scarcity_distribution` | 同 PR-K 同名性质（实测分布 + 门值即所施之门 + render） | `tests/test_minute_diagnostics_channel.py` 各条 + `tests/test_factor_eval_disclosures.py::test_valley_ridge_summarize_*`（2 条）+ `::test_summarizer_defaults_are_the_gates_the_binding_applies` | 已覆盖 |
+| `test_summarize_ridge_coverage_counterfactual_at_the_valley_floor` | 反事实：ridge 腿按 valley floor（20）还剩几天 | `tests/test_factor_eval_disclosures.py::test_valley_ridge_summarize_counterfactual_at_the_valley_floor` | 已搬迁 |
+| `test_summarize_ridge_coverage_handles_no_frames` | 空输入安全 | `tests/test_factor_eval_disclosures.py::test_valley_ridge_summarize_handles_no_frames` | 已搬迁 |
+| `test_minute_loader_blocks_when_nothing_cached` | 同 §一 #2 | 同 §一 #2 | 已覆盖 |
+| `test_evaluate_two_runs_produces_full_reports_and_incremental_axis` | 同 §一 #3 | 同 §一 #3 | 已覆盖 |
+| `test_no_book_run_never_reaches_adopt` | 同 §一 #4 | 同 §一 #4 | 已覆盖 |
+
+## 四、`tests/test_eval_valley_price_quantile_runner.py`（PR-L，14 条）
+
+| 旧 node-ID | 性质 | 新套件映射 | 状态 |
+|---|---|---|---|
+| `test_minute_loader_is_cache_only_and_discloses_empty` | 同 §一 #1 + 手算 raw 日分位（三票 0.25/0.50/0.75）+ 无效日（hi==lo）不产值 | §一 #1 各条 + `tests/test_valley_price_quantile_factor.py`（手算族：`test_hand_value_*`、`test_flat_day_with_hi_equal_lo_is_invalid`）+ `tests/test_minute_binding_vpq.py::test_service_read_through_reproduces_the_legacy_runner_cell_for_cell`（service 路径逐格复现旧 runner） | 已覆盖 |
+| `test_loader_ships_the_residual_not_the_raw_quantile` |  shipped 因子是反转中性化**残差**不是 raw 分位（防"算了又丢"接线 bug）；带截距 OLS 残差和≈0 | `tests/test_valley_price_quantile_factor.py::test_factor_is_residualized_not_raw_qbar` + `tests/test_minute_binding_vpq.py::test_combine_preserves_the_intermediate_rows_and_is_a_per_date_reduction` | 已覆盖 |
+| `test_loader_actually_consumes_the_daily_closes` | 换日频 close 面板必须改变因子值（否则上面全在 raw 分位上空过） | `tests/test_minute_binding_vpq.py::test_shifted_reversal_tracks_a_perturbed_close_bit_identically`（扰动 close → 残差逐位跟动）+ `::test_service_refuses_a_missing_daily_provider_for_vpq`（缺日频 provider loud 拒绝） | 已覆盖 |
+| `test_loader_reports_the_neutralization_coverage` | 中性化覆盖率是**实测**的（raw/rev/residual 行数、截面 min/max、raw×rev Spearman） | `tests/test_factor_eval_runner.py::test_valley_price_quantile_runs_end_to_end_with_neutralization_section`（用同一 fake **独立重算**期望值再比对——比旧测试更强，非重言式）+ `tests/test_factor_eval_disclosures.py::test_neutralization_summarize_*`（2 条汇总器单测）+ `::test_neutralization_render_matches_the_former_cli_inline_line`（渲染行逐字钉） | 已覆盖 |
+| `test_loader_blocks_when_nothing_cached` | 同 §一 #2 | 同 §一 #2 | 已覆盖 |
+| `test_thin_cross_section_leaves_the_factor_nan` | 截面 < `min_cross_section` → 该日全 NaN，绝不退回 raw | `tests/test_valley_price_quantile_factor.py::test_residualization_below_min_cross_section_is_all_nan` + `tests/test_factor_store_universe.py::test_min_cross_section_gate_still_bites_on_the_read_path` | 已覆盖 |
+| `test_summarize_neutralization_counts_missing_reversal_rows` | 汇总器对缺 reversal 行的计数 | `tests/test_factor_eval_disclosures.py::test_neutralization_summarize_counts_missing_reversal_rows` | 已搬迁 |
+| `test_summarize_neutralization_handles_an_all_missing_reversal` | 全缺 reversal → 零覆盖率、Spearman NaN | `tests/test_factor_eval_disclosures.py::test_neutralization_summarize_handles_an_all_missing_reversal` | 已搬迁 |
+| `test_reversal_from_a_daily_panel_is_the_t_minus_1_ratio` | 反转输入 = 日频面板 close 的 **T-1** 比值（不用当日 close） | `tests/test_valley_price_quantile_factor.py::test_reversal_20_is_minus_the_t_minus_1_twenty_day_return`、`::test_reversal_uses_t_minus_1_close_not_day_d` + `tests/test_minute_binding_vpq.py::test_shifted_panel_reversal_is_bit_identical_to_internal_t1`（平移面板第三条路逐比特等价钉） | 已覆盖 |
+| `test_evaluate_two_runs_produces_full_reports_and_incremental_axis` | 同 §一 #3 | 同 §一 #3 | 已覆盖 |
+| `test_extract_metrics_surfaces_the_pr_l_comparison_quantities` | 同 PR-K 同名性质 + `ic_pearson_mean` 与 rank IC 并列 + sign=+1 时 aligned spread 不误标 | 同 PR-K：性质层量在契约/标准层有钉；`_extract_metrics` 键集形状不搬迁 | 载体退役（理由 R1，见 §四） |
+| `test_spec_sign_is_positive_so_aligned_spread_is_not_mis_signed` | PR-L 的预注册 sign=+1（v0.8 成本符号缺陷只咬 sign=−1） | `tests/test_factor_requires_spec_v1.py::test_shipped_factor_declarations_match_the_d0_table[...]`（18 因子声明含 `expected_ic_sign` 参数化钉死）+ `tests/test_valley_price_quantile_factor.py::test_spec_declares_the_pre_registered_sign_and_the_pinned_deviations`；v0.8 对齐价差语义由契约层 verdict 测试钉（`tests/test_factor_eval_contract.py`，该层重构零改动） | 已覆盖 |
+| `test_no_book_run_never_reaches_adopt` | 同 §一 #4 | 同 §一 #4 | 已覆盖 |
+| `test_min_valley_bars_default_matches_the_pr_i_floor` | PR-L 复用 PR-I 的 valley floor，两 run 覆盖率可比 | **双侧钉值**：`tests/test_valley_price_quantile_factor.py::test_definition_constants_are_the_pinned_values`（`VALLEY_QUANTILE_MIN_VALLEY_BARS == 20`）+ `tests/test_valley_relative_vwap_factor.py`（`VALLEY_VWAP_MIN_VALLEY_BARS == 20`，line 401 断言）。任一侧漂移都会红其一；跨模块相等性断言语句本身不再存在 | 已覆盖（注：相等性由"同值双钉"承载，非单条等式断言） |
+
+## 五、`tests/test_eval_intraday_amp_cut_runner.py`（PR-G，4 条）
+
+| 旧 node-ID | 性质 | 新套件映射 | 状态 |
+|---|---|---|---|
+| `test_amp_cut_loader_is_cache_only_and_combines_cross_section` | cache-only + per-symbol 统计量 → 截面 z-score combine（截面归零）；stats 面板行数；empty 披露 | §一 #1 各条 + `tests/test_intraday_amp_cut_factor.py::test_amp_cut_cross_section_zscore_hand_value`、`::test_amp_cut_cross_section_below_min_is_all_nan`、`::test_amp_cut_full_pipeline_cross_section` + **D4c 中间量架构**：`tests/test_factor_store_universe.py::test_only_the_cross_sectional_factor_stores_an_intermediate`、`::test_the_cross_sectional_combine_runs_once_per_request`、`::test_served_panel_matches_the_direct_engine_and_adds_only_nan_rows[fid]` | 已覆盖 |
+| `test_amp_cut_loader_blocks_when_nothing_cached` | 同 §一 #2 | 同 §一 #2 | 已覆盖 |
+| `test_amp_cut_evaluate_two_runs_produces_full_reports_and_incremental_axis` | 同 §一 #3 | 同 §一 #3 | 已覆盖 |
+| `test_amp_cut_no_book_run_never_reaches_adopt` | 同 §一 #4 | 同 §一 #4 | 已覆盖 |
+
+## 六、PR-M（`peak_ridge_amount_ratio`）——旧套件 **0 条**（编目 BUG 3）
+
+`qt/eval_peak_ridge_amount_ratio.py` 从未有 runner 测试文件（`grep -rln
+eval_peak_ridge_amount_ratio tests/` 无命中）。新套件在此是**净增**：
+
+- `tests/test_factor_eval_disclosures.py::test_peak_summarize_counterfactual_at_the_ridge_floor`
+  与 `::test_peak_summarize_handles_no_frames`（**NET-NEW**：三个兄弟因子都有的
+  反事实 + 空帧对，PR-M 补平）。
+- `tests/test_factor_eval_disclosures.py::test_disclosure_binding_covers_exactly_the_three_publishing_factors`（披露绑定恰好覆盖 PR-J/K/M 三因子，其余因子**陈述式**无披露）。
+- 因子数学由 `tests/test_peak_ridge_amount_ratio_factor.py`（43 条）承担。
+
+## 七、新套件超出旧套件的 NET-NEW（选列，非穷举）
+
+统一 runner 落地新增、旧 59 条从未覆盖的性质：
+
+- **runner 配置门**（编目 C1/C2 collapse）：`test_factor_eval_runner.py::test_preconditions_fail_readably`（4 参数化）、`::test_missing_oos_is_a_readable_error`、`::test_invalid_book_mode_is_rejected_before_any_work`。
+- **BUG 5 闭合**：`::test_config_book_empty_or_exact_is_accepted`、`::test_config_book_mismatch_is_a_readable_error`（runner 不读 config `factors:` → 声明不符即 loud）。
+- **exec 身份 + book 双模式**：`::test_build_eval_config_declares_the_exec_identity_and_shared_kwargs`、`::test_end_to_end_wiring_both_book_modes`（含 `_bookclose` 后缀隔离）、`tests/test_factor_eval_reconcile.py` 的 `check_new_pair_consistency`（no_book/with_book 的 eval_config 只差 `book_view`）。
+- **add-Section 桥**：`test_factor_eval_disclosures.py::test_extra_section_never_moves_the_verdict_or_mandatory_sections`、`::test_an_extra_section_may_never_shadow_a_mandatory_name`、`::test_exec_basis_augmentation_seam_preserves_report_and_verdict`。
+- **诊断通道**（`test_minute_diagnostics_channel.py` 全文件，10 条）。
+- **对账 harness**（`test_factor_eval_reconcile.py`，80 条：硬闸门、panels 具名类正反向、reports 白名单、anchors）。
+- **service/store universe 语义**（`test_factor_service.py` / `test_factor_store_universe.py`：单点填≡批量填、零输出否决、(date,symbol) 缺口、中间量 universe 无关、max 形 floor 红测试等）。
+- **vpq 绑定等价**（`test_minute_binding_vpq.py`：平移面板第三条路双侧钉）。
+
+## 八、计数对照（R16 非降证据）
+
+| 桶 | 条数 |
+|---|---|
+| 旧套件总条数（10 文件） | **59** |
+| 已搬迁（逐字移入新文件） | 7（PR-J×2、PR-K×3、PR-L×2 汇总器测试） |
+| 已覆盖（性质在新套件有钉） | 50 |
+| 载体退役（性质仍在新层有钉，字面断言形态不搬迁） | 2（PR-K/PR-L 的 `extract_metrics` 键集形状，理由 R1 见下） |
+| **有映射合计** | **59 / 59（100%）** |
+| 完全无映射、无理由退役 | **0** |
+| 新套件 NET-NEW（超出旧套件的性质，选列见 §七） | ≥40（含 PR-M 补平 3 条） |
+
+**非降结论**：旧 59 条逐条有映射（7 搬迁 + 50 覆盖 + 2 载体退役但性质有钉），无一条
+静默丢失；新套件另有 NET-NEW。性质覆盖数严格上升。
+
+### 退役理由
+
+- **R1（PR-K `test_extract_metrics_surfaces_the_pr_k_comparison_quantities` / PR-L
+  `test_extract_metrics_surfaces_the_pr_l_comparison_quantities` 的键集形状）**：旧 runner
+  各自私有 `extract_metrics`，键集三档（编目 §四实测"与因子无关"）。统一 runner 把
+  metrics 收敛为单一公共子集（`qt/exec_basis_eval.py::_extract_metrics`），比较量
+  （turnover / net-by-cost 三档 / autocorr / half-life / 截面规模 / monotonicity /
+  ic_pearson）仍由评估核产出、钉在契约/标准层测试，并全文落在报告 JSON 的 section
+  payload 里——**信息无损失，损失的是"从 metrics dict 里读"这个便利载体**。该载体随
+  11 个旧 runner 在 C6 删除；若后续需要同样的 CLI 头对头摘要，应作为一个显式需求
+  重新提出，而不是靠保留旧 runner 来保住它。
+
+## 九、node-ID 抽查记录
+
+按完成定义抽查 10 条（`pytest --collect-only -q <file>` 后 grep node-ID），全命中：
+
+| # | node-ID | 结果 |
+|---|---|---|
+| 1 | `tests/test_factor_eval_providers.py::test_minute_bars_reads_normalized_bars_with_zero_live_calls` | ✅ |
+| 2 | `tests/test_factor_service.py::test_zero_output_symbol_veto_mutation` | ✅ |
+| 3 | `tests/test_factor_eval_contract.py::test_exploratory_run_is_capped_at_watch_instead_of_adopt` | ✅ |
+| 4 | `tests/test_factor_eval_standard.py::test_all_eight_sections_are_present_and_ordered` | ✅ |
+| 5 | `tests/test_factor_eval_disclosures.py::test_ridge_return_summarize_reports_the_return_guard_attrition` | ✅ |
+| 6 | `tests/test_factor_eval_runner.py::test_valley_price_quantile_runs_end_to_end_with_neutralization_section` | ✅ |
+| 7 | `tests/test_minute_binding_vpq.py::test_service_read_through_reproduces_the_legacy_runner_cell_for_cell` | ✅ |
+| 8 | `tests/test_factor_store_universe.py::test_min_cross_section_gate_still_bites_on_the_read_path` | ✅ |
+| 9 | `tests/test_exec_basis_eval.py::test_exec_basis_cli_line_survives_an_absent_metric` | ✅ |
+| 10 | `tests/test_factor_requires_spec_v1.py::test_shipped_factor_declarations_match_the_d0_table` | ✅ |

From 39e7f9c7b44f993cbf1029f21d30425950c2dc57 Mon Sep 17 00:00:00 2001
From: shaofl <2899218482@qq.com>
Date: Tue, 28 Jul 2026 02:27:42 -0700
Subject: [PATCH 3/3] docs(eval): state the index-uniqueness premise of the
 warmup cell ceiling (C5 prep review NIT)

---
 qt/factor_eval_reconcile.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/qt/factor_eval_reconcile.py b/qt/factor_eval_reconcile.py
index 0ea507d..770bb9d 100644
--- a/qt/factor_eval_reconcile.py
+++ b/qt/factor_eval_reconcile.py
@@ -795,7 +795,11 @@ def classify_panel_differences(
     ``new_only_finite`` direction explicitly requires frozen-symbol
     membership), so the class cannot exceed the ceiling by construction —
     an overshoot means the boundary/counting logic itself is broken and
-    fails the mode. AMPLITUDE inside the region stays UNBOUNDED by the
+    fails the mode. The "by construction" argument assumes (date, symbol)
+    index uniqueness on both sides; duplicate frozen rows overcount, and
+    the class then exceeds the ceiling and fails CLOSED (the safe
+    direction — bad data is never waved through). AMPLITUDE inside the
+    region stays UNBOUNDED by the
     standing ruling: for C5's geometry-reconciliation purpose that is
     acceptable, because the two loading geometries legitimately differ on
     exactly those cells (the old runner's partial pool vs the saturated