feat(nmNum): 新增 LM 高度与形状预调整的三阶段拟合

- 优先根据压力与压力导数的上下偏差调整渗透率，对齐曲线高度
- 在固定对数时间网格计算双曲线斜率残差，优先改善形状并限制整体偏离
- 按形状改善与剩余预算切换阶段，第三阶段沿用仅按整体误差接受候选的 LM
- 联合缓存数值与形状灵敏度，复用阶段间模型并同步已接受参数及曲线
- 迭代候选求解失败后交由信赖域缩步，避免相同参数重复重试
- 补充三阶段拟合日志、跟踪诊断字段及中文翻译

Bin/Config/Lang/cn/nmNum_cn.ts

@@ -664,6 +664,34 @@ Reason: %1</source>
 </context>
 <context>
     <name>nmCalculationAutoFitLM</name>
+    <message>
+        <source>Failed to refresh the sampling layer from the current curve.</source>
+        <translation>无法根据当前曲线刷新采样层级。</translation>
+    </message>
+    <message>
+        <source>LM sampling refined: %1 / %2 target points; full-target error: %3</source>
+        <translation>LM 采样加密：%1 / %2 个目标点；全目标点误差：%3</translation>
+    </message>
+    <message>
+        <source>LM sampling: layered target points (%1 / %2); acceptance uses all valid target points.</source>
+        <translation>LM 采样方式：目标点分层采样（%1 / %2 点），使用全部有效目标点验收。</translation>
+    </message>
+    <message>
+        <source>LM sampling: fixed 80 points (original mode).</source>
+        <translation>LM 采样方式：固定 80 点（原模式）。</translation>
+    </message>
+    <message>
+        <source>Full-target comparison error: initial=%1; final=%2</source>
+        <translation>全目标点对比误差：初始=%1；最终=%2</translation>
+    </message>
+    <message>
+        <source>Unavailable</source>
+        <translation>不可用</translation>
+    </message>
+    <message>
+        <source>Budget exhausted before the full sampling layer; convergence is not confirmed.</source>
+        <translation>预算已耗尽，尚未进入完整采样层，未确认收敛。</translation>
+    </message>
     <message>
         <source>=== User Stop Request Received ===</source>
         <translation>=== 用户停止请求已接收 ===</translation>
@@ -949,8 +977,8 @@ Reason: %1</source>
         <translation>连续 %1 次无有效改善，正在重建灵敏度模型进行确认</translation>
     </message>
     <message>
-        <source>Effective improvement threshold: max(%1, %2% of baseline error); %3 consecutive ineffective steps trigger convergence confirmation</source>
+        <source>Total-stage effective improvement threshold: max(%1, %2% of baseline error); %3 consecutive ineffective steps trigger convergence confirmation</source>
-        <translation>有效改善阈值：取 %1 与基准误差的 %2% 中较大值；连续 %3 次无有效改善后进行收敛确认</translation>
+        <translation>整体阶段有效改善阈值：取 %1 与基准误差的 %2% 中较大值；连续 %3 次无有效改善后进行收敛确认</translation>
     </message>
     <message>
         <source>Sensitivity probe accepted: error reduced to %1</source>
@@ -1088,6 +1116,74 @@ Reason: %1</source>
         <source>Target well name is empty</source>
         <translation>目标井名称为空</translation>
     </message>
+    <message>
+        <source>LM stage 1: align curve height using permeability only (up to %1 evaluations).</source>
+        <translation>LM 阶段一：仅调整渗透率对齐曲线高度（最多求解 %1 次）。</translation>
+    </message>
+    <message>
+        <source>Permeability alignment: k=%1, height error=%2, shape error=%3, result=%4</source>
+        <translation>渗透率对齐：k=%1，上下误差=%2，形状误差=%3，结果=%4</translation>
+    </message>
+    <message>
+        <source>LM stage 2: optimize pressure and derivative shape; stop after 3 ineffective steps.</source>
+        <translation>LM 阶段二：优先调整压力和压力导数形状，连续 3 步无明显改善后切换。</translation>
+    </message>
+    <message>
+        <source>LM stage 3: original LM fitting; accept by total error only.</source>
+        <translation>LM 阶段三：按原有 LM 拟合，仅依据整体误差接受调整。</translation>
+    </message>
+    <message>
+        <source>Sensitivity probe accepted: total error=%1</source>
+        <translation>采用灵敏度试算点：整体误差=%1</translation>
+    </message>
+    <message>
+        <source>Shape stage ended: %1</source>
+        <translation>形状阶段结束：%1</translation>
+    </message>
+    <message>
+        <source>3 consecutive steps without effective shape improvement</source>
+        <translation>连续 3 步形状没有明显改善</translation>
+    </message>
+    <message>
+        <source>reserve remaining iterations and evaluations for total fitting</source>
+        <translation>为整体拟合保留剩余迭代和求解预算</translation>
+    </message>
+    <message>
+        <source>no valid shape sensitivity model</source>
+        <translation>没有有效的形状灵敏度模型</translation>
+    </message>
+    <message>
+        <source>no feasible shape descent step</source>
+        <translation>没有满足约束的形状下降步</translation>
+    </message>
+    <message>
+        <source>Permeability alignment ended: %1</source>
+        <translation>渗透率高度对齐结束：%1</translation>
+    </message>
+    <message>
+        <source>height evaluation budget reached</source>
+        <translation>达到高度调整求解预算</translation>
+    </message>
+    <message>
+        <source>stopped by user</source>
+        <translation>用户停止</translation>
+    </message>
+    <message>
+        <source>pressure and derivative height directions conflict</source>
+        <translation>压力与压力导数的上下调整方向冲突</translation>
+    </message>
+    <message>
+        <source>curve height is approximately aligned</source>
+        <translation>曲线高度已大致对齐</translation>
+    </message>
+    <message>
+        <source>2 consecutive steps without effective height improvement</source>
+        <translation>连续 2 步上下偏差没有明显改善</translation>
+    </message>
+    <message>
+        <source>permeability reached its bound</source>
+        <translation>渗透率已到达范围边界</translation>
+    </message>
 </context>
 <context>
     <name>nmCalculationSolver</name>
@@ -3794,6 +3890,22 @@ Supported types: Vertical, Vertical Fractured, and Horizontal Multi-Fractured We
 </context>
 <context>
     <name>nmWxAutomaticFitting</name>
+    <message>
+        <source>LM sampling:</source>
+        <translation>LM 采样方式：</translation>
+    </message>
+    <message>
+        <source>Fixed sampling (80 points)</source>
+        <translation>固定采样（80 点）</translation>
+    </message>
+    <message>
+        <source>Layered target sampling</source>
+        <translation>目标点分层采样</translation>
+    </message>
+    <message>
+        <source>Layered sampling uses target-point subsets for directions and all valid target points for acceptance.</source>
+        <translation>逐层增加目标曲线采样点，并始终使用全部有效目标点判断拟合是否改善。</translation>
+    </message>
     <message>
         <source>Automatic fitting</source>
         <translation>自动拟合</translation>

Include/nmNum/nmCalculation/nmCalculationAutoFitLM.h

@@ -12,7 +12,7 @@
 #include "nmCalculation_global.h"
 
 // 固定对数时间窗口的误差诊断。时间边界为不含重叠区的基础边界；
-// rmsError 是局部加权均方根，energy 是对 total 平方的贡献。
+// rmsError 是局部加权均方根，energy 是对当前层残差能量的贡献。
 struct AutoFitTimeWindowLM {
 	double timeMin;
 	double timeMax;
@@ -25,14 +25,19 @@ struct AutoFitTimeWindowLM {
 	{}
 };
 
-// 双对数曲线误差分解。total 是 LM 候选接受和排序的唯一依据，
+// 双对数曲线误差分解。预调整使用纵向偏差，形状阶段使用双曲线斜率残差，
-// 时间窗口和残差用于 Fisher 选参，其余诊断量用于解释曲线失配。
+// 整体阶段只使用 total 接受候选；时间窗口和残差用于 Fisher 选参。
 struct AutoFitObjectiveBreakdownLM {
 	bool valid;
 	double total;
 	double pressureLoss;
 	double derivativeLoss;
 	QVector<double> residualVector;
+	// 分层模式的残差属于当前层，total 属于全部目标点；固定模式坐标数组为空。
+	QVector<double> sampleCoordinates;
+	int samplingStride;
+	int fullPointCount;
+	double layerError;
 	QVector<AutoFitTimeWindowLM> timeWindows;
 	double verticalCommonBias;
 	double verticalLoss;
@@ -41,6 +46,10 @@ struct AutoFitObjectiveBreakdownLM {
 	double horizontalLoss;
 	bool horizontalReliable;
 	bool registrationAmbiguous;
+	// 固定 log-time 网格上的双曲线斜率残差，独立于数值残差的采样层级。
+	QVector<double> shapeResiduals;
+	double pressureVerticalBias;
+	double derivativeVerticalBias;
 	double shapeLoss;
 	double lateDerivativeSlopeBias;
 	double lateDerivativeTrendLoss;
@@ -51,6 +60,9 @@ struct AutoFitObjectiveBreakdownLM {
 		, total(1.0e10)
 		, pressureLoss(std::numeric_limits<double>::quiet_NaN())
 		, derivativeLoss(std::numeric_limits<double>::quiet_NaN())
+		, samplingStride(1)
+		, fullPointCount(80)
+		, layerError(1.0e10)
 		, verticalCommonBias(std::numeric_limits<double>::quiet_NaN())
 		, verticalLoss(std::numeric_limits<double>::quiet_NaN())
 		, verticalReliable(false)
@@ -58,6 +70,8 @@ struct AutoFitObjectiveBreakdownLM {
 		, horizontalLoss(std::numeric_limits<double>::quiet_NaN())
 		, horizontalReliable(false)
 		, registrationAmbiguous(false)
+		, pressureVerticalBias(0.0)
+		, derivativeVerticalBias(0.0)
 		, shapeLoss(std::numeric_limits<double>::quiet_NaN())
 		, lateDerivativeSlopeBias(std::numeric_limits<double>::quiet_NaN())
 		, lateDerivativeTrendLoss(std::numeric_limits<double>::quiet_NaN())
@@ -97,6 +111,7 @@ public:
 	int getTotalEvaluations() const;
 	void resetOptimizer();
 	void setTargetWellName(const QString& wellName);
+	void setLayeredSamplingEnabled(bool enabled);
 
 signals:
 	void progressUpdated(int iteration, double bestFitness);
@@ -125,7 +140,7 @@ private:
 		AutoFitObjectiveBreakdownLM* breakdown,
 		QVector<QVector<double> >* curve,
 		int* elapsedMs);
-	double evaluateFitness(const QVector<double>& parameters);
+	double evaluateFitness(const QVector<double>& parameters, bool retrySolver = true);
 
 	void applyParametersToDataManager(const QVector<double>& parameters);
 	void updateReservoirParameters(const QVector<double>& parameters);
@@ -191,6 +206,9 @@ private:
 	double m_comparisonTimeMin;
 	double m_comparisonTimeMax;
 	QString m_targetWellName;
+	// 此开关由拟合界面传入；每轮重置层级，固定模式仍使用原来的 80 点目标。
+	bool m_layeredSampling;
+	int m_samplingStride;
 
 	int m_maxIterations;
 	double m_targetError;

Include/nmNum/nmSubWxs/nmWxAutomaticFitting.h

@@ -82,6 +82,8 @@ private:
 	QLineEdit* m_errorLimitEdit;
 	QComboBox* m_targetWellCombo;
 	QComboBox* m_algorithmCombo;
+	QLabel* m_samplingLabel;
+	QComboBox* m_samplingCombo;
 	QLabel* m_surrogateLabel;
 	QComboBox* m_surrogateCombo;
 

Src/nmNum/nmSubWxs/nmWxAutomaticFitting.cpp

@@ -724,6 +724,13 @@ void nmWxAutomaticFitting::setupControlPanel()
 	m_surrogateCombo->setCurrentIndex(automaticFittingData.getSurrogateScreeningEnabled() ? 1 : 0);
 	m_surrogateCombo->setMaximumWidth(160);
 	m_surrogateCombo->setMinimumWidth(160);
+	// 分层采样用于 LM 对比试验，每次打开默认使用原来的固定 80 点模式。
+	m_samplingLabel = new QLabel(tr("LM sampling:"));
+	m_samplingCombo = new QComboBox();
+	m_samplingCombo->addItem(tr("Fixed sampling (80 points)"));
+	m_samplingCombo->addItem(tr("Layered target sampling"));
+	m_samplingCombo->setMinimumWidth(180);
+	m_samplingCombo->setToolTip(tr("Layered sampling uses target-point subsets for directions and all valid target points for acceptance."));
 	connect(m_algorithmCombo, SIGNAL(currentIndexChanged(int)), this, SLOT(onAlgorithmChanged(int)));
 	onAlgorithmChanged(m_algorithmCombo->currentIndex());
 
@@ -781,6 +788,10 @@ void nmWxAutomaticFitting::setupControlPanel()
 	m_controlLayout->addWidget(m_surrogateCombo);
 	m_controlLayout->addSpacing(15);
 
+	m_controlLayout->addWidget(m_samplingLabel);
+	m_controlLayout->addWidget(m_samplingCombo);
+	m_controlLayout->addSpacing(15);
+
 	m_controlLayout->addWidget(iterationLabel);
 	m_controlLayout->addWidget(m_iterationEdit);
 	m_controlLayout->addSpacing(15);
@@ -802,6 +813,8 @@ void nmWxAutomaticFitting::onAlgorithmChanged(int index)
 	const bool usePSO = (index == 0);
 	m_surrogateLabel->setEnabled(usePSO);
 	m_surrogateCombo->setEnabled(usePSO);
+	m_samplingLabel->setEnabled(!usePSO);
+	m_samplingCombo->setEnabled(!usePSO);
 }
 
 void nmWxAutomaticFitting::setupButtons()
@@ -1189,6 +1202,7 @@ void nmWxAutomaticFitting::startAutoFitting(const QVector<QVector<double>>& targ
 		m_autoFitterLM = new nmCalculationAutoFitLM(this);
 		m_autoFitterLM->setTargetLogLogData(targetData);
 		m_autoFitterLM->setTargetWellName(targetWellName);
+		m_autoFitterLM->setLayeredSamplingEnabled(m_samplingCombo->currentIndex() == 1);
 	} else {
 		DEBUG_UI("Creating PSO auto fitter");
 		m_autoFitterPSO = new nmCalculationAutoFitPSO(this);