feat: close continuous v1 diagnostics

.gitignore

@@ -1,5 +1,8 @@
 # https://github.com/github/gitignore/blob/master/Python.gitignore
 __pycache__/
+.venv/
+data/courage_strict_v1/
+artifacts/courage_strict_v1/
 
 *.pyc
 *.pyd

AGENTS.md

@@ -0,0 +1,40 @@
+# Rolexl Qlib fork: Courage Strict V1
+
+This repository is the only active implementation workspace for `courage_strict_v1`.
+The upstream framework remote is `https://github.com/microsoft/qlib.git`; the maintained fork remote is
+`https://github.com/Rolexl/qlib.git`.
+
+## Active route
+
+- evidence route: `docs/courage_strict_v1/evidence/COURAGE_TECHNICAL_ROUTE_STRICT_20260818.md`;
+- evidence SHA-256: `ef6534868cd42e554f6dc7903e35459b42ecd21c020cf83e7340ab81938d9a91`;
+- implementation decisions: `examples/courage_strict_v1/configs/courage_strict_c1_implementation_decisions_v1.json`;
+- implementation-decision SHA-256: `baf373d37918b9ef91efd026e25e9f55bb720b8749ce698c4de0c11de651fe45`.
+
+This is the new V1 implementation. Its canonical machine identity is exactly `courage_strict_v1`; do not append
+`reproduction`, `migration`, `ACOT`, or another project qualifier. Historical Courage evidence and later engineering
+decisions remain provenance only and must not become runtime dependencies.
+
+## Storage
+
+- external source data: `/data1/lxl/workspace/datasets/.tmp/dataset`;
+- Xianyu source: `/data1/lxl/workspace/datasets/.tmp/dataset/xianyu`;
+- governed generated data: `/data1/lxl/workspace/datasets/.tmp/qlib/data/courage_strict_v1`;
+- experiment artifacts: `/data1/lxl/workspace/datasets/.tmp/qlib/artifacts/courage_strict_v1`;
+- project virtual environment: `/data1/lxl/workspace/datasets/.tmp/qlib/.venv`.
+
+No active code, configuration, catalog, command or generated artifact may resolve an ACOT path. The external raw
+vendor dataset remains read-only at `/data1/lxl/workspace/datasets/.tmp/dataset/xianyu`; source facts imported during
+the one-time transfer must be copied under this repository's `data/courage_strict_v1/source/` identity.
+
+Do not commit raw data, generated Qlib bins, checkpoints, predictions, or credentials. Never create or use paths
+named `AQuantLab` or `AQuantLab_raw`.
+
+## Project boundary
+
+Only Courage Strict V1 evidence, contracts, PIT/data kernels, V1 features/labels, Qlib dataset/model integration,
+tests, and new-run reports belong here. Do not import V2, V3, Long Alpha LA0--LA10, old authorizations, old
+scalers, old checkpoints, optimizer state, predictions, strategy, backtest, or trading code.
+
+All learned state must be fitted from scratch by this Qlib V1. April 2026 is Development replay only;
+May is historical-consumed evidence and June remains sealed. ACC is reported as a diagnostic, not a selection gate.

docs/TRAINING_ARTIFACT_POLICY.md

@@ -0,0 +1,46 @@
+# 每次训练的 Train/Valid Loss 曲线强制规则
+
+状态：`ACCEPTED_REPOSITORY_WIDE_TRAINING_RULE`。
+
+本规则适用于此仓库之后的每一次正式模型训练，不限于某个 Courage 版本。参考图为 [Courage V1 step curve](courage_strict_v1_step_curve/TRAIN_VALID_LOSS_CURVE.png)。
+
+## 必须记录
+
+- 在配置冻结的每个 validation step 同时记录当前 Train window loss 与完整固定 Valid loss。
+- Train loss 必须是刚结束窗口的损失，不能用从step 1开始的累计平均值冒充局部训练曲线。
+- Valid population、Label mask、scaler和选择loss在一次运行中必须固定。
+- `loss_curve.json` 必须保留原始逐点值、global step、学习率、样本覆盖和best checkpoint选择指标。
+
+## 必须绘图
+
+每次训练必须生成：
+
+1. 运行目录中的 `TRAIN_VALID_LOSS_CURVE.png`；
+2. 训练报告文档目录中的同名字节副本；
+3. 图中使用global optimizer step作为横轴，绘制未平滑的Train与Valid loss，并标出best step；
+4. 运行目录和训练报告目录中的 `PER_HEAD_TRAIN_VALID_LOSS_CURVES.png`，分别展示每个horizon的Train-window与完整固定Valid曲线；
+5. 训练报告必须链接两张图，并写明验证间隔、总验证点数和best step。
+
+gradient/LR曲线建议同时生成，但不能替代上述两张Train/Valid曲线。
+
+## 最佳checkpoint分组诊断
+
+日期、分钟位置、PIT行业和换手率分组不在每个validation step重复计算。训练完成并冻结best checkpoint后，必须对完整固定Valid自动执行一次，生成`BEST_CHECKPOINT_GROUPED_METRICS.csv`或内容等价且被manifest固定的产物。
+
+- 分组至少覆盖`date`、`minute_slot`、`industry_id`、`turnover_band`；
+- 必须覆盖所有已训练horizon，并保存样本数、RMSE、bias、ACC、BAcc、MCC、AUC、target/prediction up rate；
+- 该步骤只允许读取训练时已经授权的固定Valid，不能用分组结果重新选择checkpoint；
+- 若基线门要求先比较zero、Train mean/median、momentum/reversal和Ridge，则分组诊断应在同一post-training closure中完成。
+
+## 完成门禁
+
+以下任一情况都标记为 `TRAINING_ARTIFACTS_INCOMPLETE`：
+
+- 缺少`loss_curve.json`或`TRAIN_VALID_LOSS_CURVE.png`；
+- 缺少逐horizon Train/Valid曲线或best-checkpoint分组指标；
+- 图与JSON的step或best checkpoint不一致；
+- 文档副本缺失或与运行目录不是相同字节；
+- 只有epoch末值，无法观察epoch内部变化；
+- 用平滑值替换原始loss，或者Train和Valid使用不同选择口径。
+
+该门禁只约束训练可观测性，不授权读取新数据、训练、refit、策略、回测或交易。

docs/courage_strict_continuous_v1/BASELINE_CLOSURE_EVALUATION.md

@@ -0,0 +1,162 @@
+# Courage Strict Continuous V1 测评基线闭环
+
+> 冻结 step-2250 checkpoint 的 Train-only baseline 与完整 January rolling Valid 诊断；不是未见 Test。
+
+## 结论
+
+- 门禁：`FAIL_BASELINE_GATE`；平均 best-baseline RMSE skill `-0.184%`，正 skill `0/7`。
+- 当前模型未改动，checkpoint 选择未重做；全部 baseline 在完全相同的 Valid key/Label mask 上计算。
+
+## 诊断解读
+
+- 模型输出明显收缩：七个 horizon 的 `prediction_std / target_std` 仅 `1.0%`～`12.4%`；5m/15m prediction up rate 分别只有 `0.054%`/`0.405%`。
+- Target 水平发生时间漂移：120m Train→Valid mean 为 `-0.022%`→`+0.075%`，240m 为 `-0.037%`→`+0.140%`，480m 为 `+0.048%`→`+0.430%`。
+- 240m/480m 仍有弱排序信息（AUC `54.180%`/`54.381%`，Rank IC `+0.06176`/`+0.05940`），但绝对收益点预测均输给常数基线；这是“排序信号存在、条件均值校准失败”，不能视为收益回归通过。
+- 最强 RMSE baseline 分布为：5/15/30m Ridge，60/120/240m zero，480m Train mean；继续增加 epoch 或更换网络不能直接解释这一模式。
+
+## 三项 Sanity Check 闭环
+
+基线闭环之后，对同一 `courage_strict_continuous_v1` 数据链路和模型实现执行了三项有界 sanity check。它们用于区分明显的数据/实现错误与真实的泛化失败，不用于重新选择或晋升 checkpoint。
+
+| 检查 | 执行口径 | 结果 | 结论 |
+|---|---|---:|---|
+| 原始 Bar Label 独立复算 | 每个 horizon 时间分层抽取15条，共105条；从不可变分钟 Parquet 以 `amount/volume` 重算 entry/exit VWAP 和收益，再与 provider float32 bin 做 bit-level 比较 | 105/105 bit-exact；最大误差0 | `PASS_LABEL_RECOMPUTATION` |
+| 固定小样本过拟合 | 固定512条七头均成熟样本；dropout=0、weight decay=0；最多3000 steps | 标准化 Huber `0.585392 → 0.00009093`，下降99.984%；七头全部过门 | `PASS_SMALL_SAMPLE_OVERFIT` |
+| Shuffled-label 随机对照 | Train 标签按 horizon 独立置换；保留输入、mask和完整 January Valid | 首轮3 seeds × 7 heads 为19/21过门，480m与15m各有一个孤立越界 | 首轮严格记为 `FAIL_SHUFFLED_LABEL_RANDOM_CONTROL` |
+
+### Shuffled-label 后续复核
+
+由于首轮两个异常没有在同一 horizon 跨 seed 重复，随后预注册并执行了更强的 null-distribution 复核：10个配对 seed；每个 seed 先评估 initialization-only，再训练500-step shuffled-label；两种条件都评估完整 January Valid，并以交易日为单位进行3日 circular moving-block bootstrap（5,000次）。
+
+升级泄漏调查必须同时满足：
+
+1. shuffled 聚合95% CI排除随机值；
+2. shuffled−initialization 配对95% CI同方向排除0；
+3. 至少4/10 seed的单独交易日 block CI同方向排除随机值。
+
+最终没有任何 horizon/metric 同时满足三项条件，判定为 `PASS_NO_REPEATABLE_SHUFFLED_LABEL_LEAKAGE_PATTERN`。原两个异常的关键复核如下：
+
+| 指标 | Shuffled mean [95% CI] | Shuffled−Init [95% CI] | 泄漏升级 |
+|---|---:|---:|---|
+| 480m AUC | `0.50940 [0.49695, 0.52223]` | `0.01166 [-0.00876, 0.03148]` | NO |
+| 480m Rank IC | `0.00857 [-0.00946, 0.02733]` | `0.01027 [-0.02200, 0.04137]` | NO |
+| 15m Rank IC | `-0.00763 [-0.01637, 0.00046]` | `-0.01665 [-0.03160, -0.00215]` | NO |
+
+综合解释：已抽查 Label 的公式与落盘值正确；模型、mask、标准化和优化器具备正常记忆能力；没有发现可重复的 shuffled-label 泄漏模式。因此“明显实现错误或稳定泄漏”从首要原因中降级，但这不改变正式模型 `FAIL_BASELINE_GATE`，下一步仍应聚焦输出收缩、绝对收益条件均值校准和时间状态漂移。
+
+详细证据：
+
+- [三项 Sanity Check 报告](SANITY_CHECK_REPORT.md)
+- [Initialization-only × Shuffled-label Null Distribution](SANITY_NULL_DISTRIBUTION_REPORT.md)
+
+## 逐 horizon 最强基线
+
+| H | Model RMSE | Best baseline | Baseline RMSE | Skill | ACC | BAcc | MCC | AUC | Rank IC | Pred/Target Std |
+|---:|---:|:--|---:|---:|---:|---:|---:|---:|---:|---:|
+| 5 | 0.00426189 | ridge_terminal | 0.00425595 | -0.140% | 51.845% | 50.006% | +0.00250 | 50.942% | +0.03486 | 0.010 |
+| 15 | 0.00762374 | ridge_terminal | 0.00761276 | -0.144% | 51.818% | 50.037% | +0.00589 | 51.628% | +0.04646 | 0.023 |
+| 30 | 0.01070863 | ridge_terminal | 0.01069625 | -0.116% | 51.536% | 50.289% | +0.01756 | 52.362% | +0.05342 | 0.026 |
+| 60 | 0.01494003 | zero | 0.01492692 | -0.088% | 51.826% | 51.207% | +0.03565 | 52.852% | +0.05751 | 0.034 |
+| 120 | 0.02058491 | zero | 0.02056740 | -0.085% | 52.171% | 52.160% | +0.04499 | 53.319% | +0.05724 | 0.070 |
+| 240 | 0.02858734 | zero | 0.02849929 | -0.309% | 52.449% | 52.755% | +0.05711 | 54.180% | +0.06176 | 0.102 |
+| 480 | 0.04235313 | train_mean | 0.04218151 | -0.407% | 53.177% | 52.997% | +0.06004 | 54.381% | +0.05940 | 0.124 |
+
+## 全部 estimator RMSE
+
+| H | Estimator | RMSE | Skill vs best baseline | BAcc | MCC | AUC | Rank IC |
+|---:|:--|---:|---:|---:|---:|---:|---:|
+| 5 | model | 0.00426189 | -0.140% | 50.006% | +0.00250 | 50.942% | +0.03486 |
+| 5 | zero | 0.00425740 | -0.034% | 50.000% | +0.00000 | 50.000% | +nan |
+| 5 | train_mean | 0.00425716 | -0.028% | 50.000% | +0.00000 | 50.000% | +nan |
+| 5 | train_median | 0.00425970 | -0.088% | 50.000% | +0.00000 | 50.000% | +nan |
+| 5 | momentum_ret5 | 0.00659167 | -54.881% | 48.844% | -0.02358 | 48.693% | -0.03358 |
+| 5 | reversal_ret5 | 0.00646818 | -51.980% | 50.853% | +0.01717 | 51.307% | +0.03358 |
+| 5 | ridge_terminal | 0.00425595 | +0.000% | 51.088% | +0.02257 | 51.396% | +0.05035 |
+| 15 | model | 0.00762374 | -0.144% | 50.037% | +0.00589 | 51.628% | +0.04646 |
+| 15 | zero | 0.00761525 | -0.033% | 50.000% | +0.00000 | 50.000% | +nan |
+| 15 | train_mean | 0.00761510 | -0.031% | 50.000% | +0.00000 | 50.000% | +nan |
+| 15 | train_median | 0.00762208 | -0.123% | 50.000% | +0.00000 | 50.000% | +nan |
+| 15 | momentum_ret5 | 0.00918254 | -20.620% | 48.389% | -0.03288 | 48.014% | -0.03202 |
+| 15 | reversal_ret5 | 0.00894093 | -17.447% | 51.372% | +0.02763 | 51.986% | +0.03202 |
+| 15 | ridge_terminal | 0.00761276 | +0.000% | 51.177% | +0.02354 | 51.528% | +0.05941 |
+| 30 | model | 0.01070863 | -0.116% | 50.289% | +0.01756 | 52.362% | +0.05342 |
+| 30 | zero | 0.01069731 | -0.010% | 50.000% | +0.00000 | 50.000% | +nan |
+| 30 | train_mean | 0.01069742 | -0.011% | 50.000% | +0.00000 | 50.000% | +nan |
+| 30 | train_median | 0.01071025 | -0.131% | 50.000% | +0.00000 | 50.000% | +nan |
+| 30 | momentum_ret5 | 0.01185302 | -10.815% | 48.863% | -0.02322 | 48.765% | -0.02358 |
+| 30 | reversal_ret5 | 0.01166981 | -9.102% | 50.760% | +0.01531 | 51.235% | +0.02358 |
+| 30 | ridge_terminal | 0.01069625 | +0.000% | 51.105% | +0.02211 | 51.531% | +0.06406 |
+| 60 | model | 0.01494003 | -0.088% | 51.207% | +0.03565 | 52.852% | +0.05751 |
+| 60 | zero | 0.01492692 | +0.000% | 50.000% | +0.00000 | 50.000% | +nan |
+| 60 | train_mean | 0.01492926 | -0.016% | 50.000% | +0.00000 | 50.000% | +nan |
+| 60 | train_median | 0.01495127 | -0.163% | 50.000% | +0.00000 | 50.000% | +nan |
+| 60 | momentum_ret5 | 0.01581712 | -5.964% | 48.835% | -0.02379 | 48.828% | -0.01629 |
+| 60 | reversal_ret5 | 0.01558638 | -4.418% | 50.570% | +0.01148 | 51.172% | +0.01629 |
+| 60 | ridge_terminal | 0.01493721 | -0.069% | 50.907% | +0.01841 | 51.739% | +0.06404 |
+| 120 | model | 0.02058491 | -0.085% | 52.160% | +0.04499 | 53.319% | +0.05724 |
+| 120 | zero | 0.02056740 | +0.000% | 50.000% | +0.00000 | 50.000% | +nan |
+| 120 | train_mean | 0.02057639 | -0.044% | 50.000% | +0.00000 | 50.000% | +nan |
+| 120 | train_median | 0.02061390 | -0.226% | 50.000% | +0.00000 | 50.000% | +nan |
+| 120 | momentum_ret5 | 0.02125616 | -3.349% | 48.777% | -0.02500 | 48.955% | -0.01372 |
+| 120 | reversal_ret5 | 0.02101251 | -2.164% | 50.286% | +0.00577 | 51.045% | +0.01372 |
+| 120 | ridge_terminal | 0.02060984 | -0.206% | 51.583% | +0.03285 | 52.780% | +0.07312 |
+| 240 | model | 0.02858734 | -0.309% | 52.755% | +0.05711 | 54.180% | +0.06176 |
+| 240 | zero | 0.02849929 | +0.000% | 50.000% | +0.00000 | 50.000% | +nan |
+| 240 | train_mean | 0.02851954 | -0.071% | 50.000% | +0.00000 | 50.000% | +nan |
+| 240 | train_median | 0.02857717 | -0.273% | 50.000% | +0.00000 | 50.000% | +nan |
+| 240 | momentum_ret5 | 0.02892367 | -1.489% | 49.439% | -0.01146 | 49.866% | -0.00517 |
+| 240 | reversal_ret5 | 0.02889172 | -1.377% | 49.676% | -0.00654 | 50.134% | +0.00517 |
+| 240 | ridge_terminal | 0.02861314 | -0.399% | 53.032% | +0.06342 | 54.443% | +0.06444 |
+| 480 | model | 0.04235313 | -0.407% | 52.997% | +0.06004 | 54.381% | +0.05940 |
+| 480 | zero | 0.04222734 | -0.109% | 50.000% | +0.00000 | 50.000% | +nan |
+| 480 | train_mean | 0.04218151 | +0.000% | 50.000% | +0.00000 | 50.000% | +nan |
+| 480 | train_median | 0.04235324 | -0.407% | 50.000% | +0.00000 | 50.000% | +nan |
+| 480 | momentum_ret5 | 0.04250367 | -0.764% | 49.538% | -0.00942 | 49.953% | -0.00054 |
+| 480 | reversal_ret5 | 0.04249819 | -0.751% | 49.663% | -0.00680 | 50.047% | +0.00054 |
+| 480 | ridge_terminal | 0.04234481 | -0.387% | 54.407% | +0.08807 | 55.747% | +0.06072 |
+
+## 逐日稳定性摘要
+
+| H | Days | IC>0 days | BAcc>50% days | Mean daily MCC |
+|---:|---:|---:|---:|---:|
+| 5 | 20 | 19 | 12 | +0.00176 |
+| 15 | 20 | 19 | 17 | +0.00614 |
+| 30 | 20 | 19 | 19 | +0.01776 |
+| 60 | 20 | 18 | 19 | +0.02985 |
+| 120 | 20 | 16 | 18 | +0.03496 |
+| 240 | 19 | 14 | 15 | +0.04083 |
+| 480 | 18 | 9 | 15 | +0.05124 |
+
+## Prediction 收缩
+
+`Pred/Target Std` 显著小于 1 表示模型输出接近常数；完整均值、标准差和分位数见 `target_prediction_distribution.csv`。
+
+## 产物
+
+- `baseline_metrics.csv`：全部 estimator × horizon 指标；
+- `daily_metrics.csv`：逐日 IC/BAcc/MCC/RMSE；
+- `grouped_model_metrics.csv`：日期、分钟位置、行业和换手率分组；
+- `feature_distribution.csv`：Train/Valid 终点特征缺失率、median/IQR；
+- `target_prediction_distribution.csv`：Target/Prediction 均值、标准差、分位数和上涨比例；
+- `baseline_predictions.parquet`：同人口 baseline 预测；
+- `ridge_coefficients.json`：Train-only Ridge 系数和拟合人口。
+- `sanity_checks_v1/`：Label逐条复算、小样本过拟合曲线与3-seed shuffled-label结果；
+- `sanity_null_distribution_v1/`：10-seed配对日级指标、bootstrap摘要与最终机器判定。
+
+## 边界
+
+- Ridge 只使用信号时点的 12 个动态特征、5 个慢特征、缺失指示和时钟变量；不使用 Valid 拟合或调参。
+- Momentum/Reversal 是可观测 `stock_ret_5` 的原值/相反数，缺失时预测 0。
+- 基线评测本身未重新训练 PatchTST；其后的三项 sanity check 使用独立诊断命名空间，未覆盖或晋升正式 checkpoint。
+- 未重训或改选正式 PatchTST；仅训练不可晋升的 sanity 诊断模型。未读取 2026-02-02 及以后数据，未执行 refit、策略、回测、交易或远端推送。
+
+## 后续 March 漂移诊断闭环（只读）
+
+基线评测完成后，另以冻结 step-2250 checkpoint 及既有 January—March 预测补齐了以下事后诊断：
+
+- March 逐日 prediction/target 均值、标准差、上涨率、BAcc、MCC 与 Rank IC；
+- 按 30 分钟位置桶、PIT 行业和 T-1 换手率档分组；
+- Train→January→February→March 的特征缺失率、median/IQR、PSI 与标签分布漂移；
+- 每期系统抽样 200,000 条序列，扫描完整 1,200 分钟内部历史位置的可用率和连续缺失长度。
+
+结果显示 March 慢特征及绝对收益基准发生明显漂移，而完整 lookback 动态特征平均可用率仍约 90%，不支持“历史窗口因大面积缺失而失效”。详细结果见 [March 漂移诊断闭环](MARCH_DRIFT_DIAGNOSTIC_CLOSURE.md)。该后续诊断读取范围严格早于 2026-04-01；未重训、改选 checkpoint 或读取 April 及以后数据。