@reicek/neataptic-ts 0.1.24 → 0.1.26

package/src/architecture/network/onnx/export/README.md

@@ -20,380 +20,380 @@ flowchart LR
 
 ## architecture/network/onnx/export/network.onnx.export.types.ts
 
-### OnnxExportOptions
+### ActivationSquashFunction
 
-Options controlling ONNX-like export.
+```ts
+ActivationSquashFunction(
+  x: number,
+  derivate: boolean | undefined,
+): number
+```
 
-These options trade off strictness, portability, and fidelity:
+Activation function signature used by ONNX layer emission helpers.
 
-- **Strict (default-ish)** export tries to keep the graph easy to interpret:
-  layered topology, homogeneous activations per layer, and fully-connected layers.
+### ConvInferenceEvaluationContext
 
-- **Relaxed** export (`allowPartialConnectivity` / `allowMixedActivations`) can represent
-  more networks, but it may generate graphs that are primarily meant for NeatapticTS’s
-  importer (and may be less friendly to external ONNX tooling).
+Width and shape evaluation context used by Conv inference helpers.
 
-- **Recurrent export** (`allowRecurrent`) is intentionally conservative and currently
-  focuses on a constrained single-step representation and optional fused heuristics.
+### ConvInferenceKernelEvaluationContext
 
-Key fields (high-level):
-- `includeMetadata`: includes `metadata_props` with architecture hints.
-- `opset`: numeric opset version stored in the exported model metadata (default is
-  resolved by the exporter; commonly 18 in this codebase).
-- `legacyNodeOrdering`: keeps older node ordering for backward compatibility.
-- `conv2dMappings` / `pool2dMappings`: encode conv/pool semantics for fully-connected
-  layers via explicit mapping declarations.
+Kernel candidate context for one Conv inference evaluation pass.
 
-### ExportNodeIndexAssignmentContext
+### ConvInferenceResult
 
-Context for assigning a stable export index to one node.
+Collected inferred Conv metadata payload.
 
-### LstmPatternStub
+### ConvInferenceTraversalContext
 
-Heuristic LSTM pattern stub for metadata output.
+Traversal context for one hidden layer during Conv inference.
 
-### LstmLayerTraversalContext
+### ConvKernelConsistencyContext
 
-Traversal context for one hidden layer during LSTM stub collection.
+Context for kernel-coordinate consistency checks at one output position.
 
-### LstmCandidateContext
+### ConvLayerPairContext
 
-Candidate context for validating one LSTM-like hidden layer pattern.
+Context for one resolved Conv mapping layer pair.
 
-### ConvInferenceTraversalContext
+### ConvOutputCoordinate
 
-Traversal context for one hidden layer during Conv inference.
+Coordinate for one Conv output neuron position.
 
-### ConvInferenceEvaluationContext
+### ConvRepresentativeKernelContext
 
-Width and shape evaluation context used by Conv inference helpers.
+Context for representative Conv kernel collection per output channel.
 
-### ConvInferenceKernelEvaluationContext
+### ConvSharingValidationContext
 
-Kernel candidate context for one Conv inference evaluation pass.
+Context for validating Conv sharing across all declared mappings.
 
-### ConvInferenceResult
+### ConvSharingValidationResult
 
-Collected inferred Conv metadata payload.
+Result of Conv sharing validation across declared mappings.
 
-### OnnxBuildResolvedOptions
+### DenseActivationContext
 
-Resolved options used by ONNX model build orchestration.
+Dense activation emission context.
 
-### OnnxGraphDimensionBuildContext
+### DenseActivationNodePayload
 
-Context for constructing input/output ONNX graph dimensions.
+Strongly typed activation node payload used by dense export helpers.
 
-### OnnxGraphDimensions
+### DenseGemmNodePayload
 
-Output dimensions used by ONNX graph input/output value info payloads.
+Strongly typed Gemm node payload used by dense export helpers.
 
-### OnnxBaseModelBuildContext
+### DenseGraphNames
 
-Context for constructing a base ONNX model shell.
+Dense graph tensor names.
 
-### OnnxModelMetadataContext
+### DenseInitializerValues
 
-Context for applying optional ONNX model metadata.
+Dense initializer value arrays.
 
-### OnnxRecurrentCollectionContext
+### DenseLayerContext
 
-Context for collecting recurrent layer indices during model build.
+Dense layer context enriched with resolved activation function.
 
-### OnnxRecurrentLayerTraversalContext
+### DenseLayerParams
 
-Traversal context for one hidden layer during recurrent-input collection.
+Parameters for dense layer emission.
 
-### OnnxRecurrentInputValueInfoContext
+### DenseOrderedNodePayload
 
-Context for constructing one recurrent previous-state graph input payload.
+Dense node payload union used by ordered append helpers.
 
-### OnnxRecurrentLayerProcessingContext
+### DenseTensorNames
 
-Execution context for processing one hidden recurrent layer.
+Dense initializer tensor names.
 
-### OnnxLayerEmissionResult
+### DenseWeightBuildContext
 
-Result of emitting non-input export layers.
+Context for building dense layer initializers from two adjacent layers.
 
-### OnnxLayerEmissionContext
+### DenseWeightBuildResult
 
-Context for emitting non-input layers during model build.
+Dense layer initializer fold output.
 
-### LayerBuildContext
+### DenseWeightRow
 
-Layer build context used while emitting one ONNX graph layer segment.
+One collected dense row before fold to flattened initializers.
 
-### LayerTraversalContext
+### DenseWeightRowCollectionContext
 
-Layer traversal context with adjacent layers and output classification.
+Context for collecting one dense row.
 
-### LayerActivationContext
+### DiagonalRecurrentBuildContext
 
-Activation analysis context for one layer.
+Context for building a diagonal recurrent matrix from self-connections.
 
-### LayerRecurrentDecisionContext
+### ExportNodeIndexAssignmentContext
 
-Context used to decide recurrent emission branch usage.
+Context for assigning a stable export index to one node.
 
-### OnnxPostProcessingContext
+### FlattenAfterPoolingContext
 
-Context for post-processing and export metadata finalization.
+Flatten emission context after optional pooling.
 
-### RecurrentHeuristicEmissionContext
+### FusedRecurrentEmissionExecutionContext
 
-Context for heuristic recurrent operator emission traversal.
+Shared execution context for emitting one fused recurrent layer payload.
 
-### HiddenLayerHeuristicContext
+### FusedRecurrentGraphNames
 
-Context for one hidden layer during heuristic recurrent emission.
+Context for ONNX fused recurrent node payload names.
 
-### LstmEmissionContext
+### FusedRecurrentInitializerNames
 
-Context for heuristic LSTM emission when a layer matches expected shape.
+Context for ONNX fused recurrent initializer names.
 
 ### GruEmissionContext
 
 Context for heuristic GRU emission when a layer matches expected shape.
 
-### RecurrentGateRowCollectionContext
+### HiddenLayerHeuristicContext
 
-Context for collecting one recurrent gate row (one neuron).
+Context for one hidden layer during heuristic recurrent emission.
 
-### RecurrentGateRow
+### IndexedMetadataAppendContext
 
-One recurrent gate row payload before flatten fold.
+Append-an-index metadata context for JSON-array metadata keys.
 
-### RecurrentGateParameterCollectionResult
+### LayerActivationContext
 
-Flattened recurrent gate parameter vectors for one fused operator.
+Activation analysis context for one layer.
 
-### RecurrentGateBlockCollectionContext
+### LayerBuildContext
 
-Context for collecting one gate parameter block.
+Layer build context used while emitting one ONNX graph layer segment.
 
-### FusedRecurrentInitializerNames
+### LayerRecurrentDecisionContext
 
-Context for ONNX fused recurrent initializer names.
+Context used to decide recurrent emission branch usage.
 
-### FusedRecurrentGraphNames
+### LayerTraversalContext
 
-Context for ONNX fused recurrent node payload names.
+Layer traversal context with adjacent layers and output classification.
 
-### FusedRecurrentEmissionExecutionContext
+### LstmCandidateContext
 
-Shared execution context for emitting one fused recurrent layer payload.
+Candidate context for validating one LSTM-like hidden layer pattern.
 
-### RecurrentLayerEmissionParams
+### LstmEmissionContext
 
-Parameters for single-step recurrent layer emission.
+Context for heuristic LSTM emission when a layer matches expected shape.
 
-### RecurrentLayerEmissionContext
+### LstmLayerTraversalContext
 
-Derived execution context for single-step recurrent layer emission.
+Traversal context for one hidden layer during LSTM stub collection.
 
-### RecurrentInitializerNames
+### LstmPatternStub
 
-Initializer tensor names for one single-step recurrent layer.
+Heuristic LSTM pattern stub for metadata output.
 
-### RecurrentInitializerValues
+### OnnxBaseModelBuildContext
 
-Collected initializer vectors for one single-step recurrent layer.
+Context for constructing a base ONNX model shell.
 
-### RecurrentInitializerEmissionContext
+### OnnxBuildResolvedOptions
 
-Context for pushing recurrent initializers into ONNX graph state.
+Resolved options used by ONNX model build orchestration.
 
-### RecurrentGemmEmissionContext
+### OnnxConvEmissionContext
 
-Context for emitting one Gemm node for recurrent single-step export.
+Context used after resolving Conv mapping for one layer.
 
-### RecurrentGraphNames
+### OnnxConvEmissionParams
 
-Derived graph names for one recurrent single-step layer payload.
+Parameters accepted by Conv layer emission.
 
-### RecurrentActivationEmissionContext
+### OnnxConvParameters
 
-Context for selecting and emitting recurrent activation node payload.
+Flattened Conv parameters for ONNX initializers.
 
-### ConvSharingValidationResult
+### OnnxConvTensorNames
 
-Result of Conv sharing validation across declared mappings.
+Tensor names generated for Conv parameters.
 
-### ConvSharingValidationContext
+### OnnxExportOptions
 
-Context for validating Conv sharing across all declared mappings.
+Options controlling ONNX-like export.
 
-### ConvLayerPairContext
+These options trade off strictness, portability, and fidelity:
 
-Context for one resolved Conv mapping layer pair.
+- **Strict (default-ish)** export tries to keep the graph easy to interpret:
+  layered topology, homogeneous activations per layer, and fully-connected layers.
 
-### ConvOutputCoordinate
+- **Relaxed** export (`allowPartialConnectivity` / `allowMixedActivations`) can represent
+  more networks, but it may generate graphs that are primarily meant for NeatapticTS’s
+  importer (and may be less friendly to external ONNX tooling).
 
-Coordinate for one Conv output neuron position.
+- **Recurrent export** (`allowRecurrent`) is intentionally conservative and currently
+  focuses on a constrained single-step representation and optional fused heuristics.
 
-### ConvRepresentativeKernelContext
+Key fields (high-level):
+- `includeMetadata`: includes `metadata_props` with architecture hints.
+- `opset`: numeric opset version stored in the exported model metadata (default is
+  resolved by the exporter; commonly 18 in this codebase).
+- `legacyNodeOrdering`: keeps older node ordering for backward compatibility.
+- `conv2dMappings` / `pool2dMappings`: encode conv/pool semantics for fully-connected
+  layers via explicit mapping declarations.
 
-Context for representative Conv kernel collection per output channel.
+### OnnxGraphDimensionBuildContext
 
-### ConvKernelConsistencyContext
+Context for constructing input/output ONNX graph dimensions.
 
-Context for kernel-coordinate consistency checks at one output position.
+### OnnxGraphDimensions
 
-### WeightToleranceComparisonContext
+Output dimensions used by ONNX graph input/output value info payloads.
 
-Context for comparing two scalar weights with numeric tolerance.
+### OnnxLayerEmissionContext
 
-### OnnxConvEmissionParams
+Context for emitting non-input layers during model build.
 
-Parameters accepted by Conv layer emission.
+### OnnxLayerEmissionResult
 
-### OnnxConvEmissionContext
+Result of emitting non-input export layers.
 
-Context used after resolving Conv mapping for one layer.
+### OnnxModelMetadataContext
 
-### OnnxConvParameters
+Context for applying optional ONNX model metadata.
 
-Flattened Conv parameters for ONNX initializers.
+### OnnxPostProcessingContext
 
-### OnnxConvTensorNames
+Context for post-processing and export metadata finalization.
 
-Tensor names generated for Conv parameters.
+### OnnxRecurrentCollectionContext
 
-### ActivationSquashFunction
+Context for collecting recurrent layer indices during model build.
 
-```ts
-ActivationSquashFunction(
-  x: number,
-  derivate: boolean | undefined,
-): number
-```
+### OnnxRecurrentInputValueInfoContext
 
-Activation function signature used by ONNX layer emission helpers.
+Context for constructing one recurrent previous-state graph input payload.
 
-### SharedGemmNodeBuildParams
+### OnnxRecurrentLayerProcessingContext
 
-Shared parameters for constructing a Gemm node payload.
+Execution context for processing one hidden recurrent layer.
 
-### SharedActivationNodeBuildParams
+### OnnxRecurrentLayerTraversalContext
 
-Shared parameters for constructing an activation node payload.
+Traversal context for one hidden layer during recurrent-input collection.
 
 ### OptionalLayerOutputParams
 
 Shared parameters for optional pooling/flatten output emission.
 
-### DenseWeightBuildContext
-
-Context for building dense layer initializers from two adjacent layers.
-
-### DenseWeightRow
+### OptionalPoolingAndFlattenParams
 
-One collected dense row before fold to flattened initializers.
+Parameters for optional pooling + flatten emission after a layer output.
 
-### DenseWeightBuildResult
+### PerNeuronConcatNodePayload
 
-Dense layer initializer fold output.
+Per-neuron concat node payload.
 
-### DenseWeightRowCollectionContext
+### PerNeuronGraphNames
 
-Context for collecting one dense row.
+Per-neuron graph tensor names.
 
-### DiagonalRecurrentBuildContext
+### PerNeuronLayerContext
 
-Context for building a diagonal recurrent matrix from self-connections.
+Per-neuron layer context alias.
 
-### RecurrentRowCollectionContext
+### PerNeuronLayerParams
 
-Context for collecting one recurrent matrix row.
+Parameters for per-neuron layer emission.
 
-### OptionalPoolingAndFlattenParams
+### PerNeuronNodeContext
 
-Parameters for optional pooling + flatten emission after a layer output.
+Per-neuron normalized node context.
 
-### PoolingEmissionContext
+### PerNeuronSubgraphContext
 
-Pooling emission context resolved for one layer output.
+Per-neuron subgraph emission context.
 
-### FlattenAfterPoolingContext
+### PerNeuronTensorNames
 
-Flatten emission context after optional pooling.
+Per-neuron initializer tensor names.
 
 ### PoolingAttributes
 
 Pooling tensor attributes for ONNX node payloads.
 
-### IndexedMetadataAppendContext
+### PoolingEmissionContext
 
-Append-an-index metadata context for JSON-array metadata keys.
+Pooling emission context resolved for one layer output.
 
-### SpecMetadataAppendContext
+### RecurrentActivationEmissionContext
 
-Append-a-spec metadata context for JSON-array metadata keys.
+Context for selecting and emitting recurrent activation node payload.
 
-### DenseLayerParams
+### RecurrentGateBlockCollectionContext
 
-Parameters for dense layer emission.
+Context for collecting one gate parameter block.
 
-### DenseLayerContext
+### RecurrentGateParameterCollectionResult
 
-Dense layer context enriched with resolved activation function.
+Flattened recurrent gate parameter vectors for one fused operator.
 
-### DenseTensorNames
+### RecurrentGateRow
 
-Dense initializer tensor names.
+One recurrent gate row payload before flatten fold.
 
-### DenseInitializerValues
+### RecurrentGateRowCollectionContext
 
-Dense initializer value arrays.
+Context for collecting one recurrent gate row (one neuron).
 
-### DenseGraphNames
+### RecurrentGemmEmissionContext
 
-Dense graph tensor names.
+Context for emitting one Gemm node for recurrent single-step export.
 
-### DenseActivationContext
+### RecurrentGraphNames
 
-Dense activation emission context.
+Derived graph names for one recurrent single-step layer payload.
 
-### DenseGemmNodePayload
+### RecurrentHeuristicEmissionContext
 
-Strongly typed Gemm node payload used by dense export helpers.
+Context for heuristic recurrent operator emission traversal.
 
-### DenseActivationNodePayload
+### RecurrentInitializerEmissionContext
 
-Strongly typed activation node payload used by dense export helpers.
+Context for pushing recurrent initializers into ONNX graph state.
 
-### DenseOrderedNodePayload
+### RecurrentInitializerNames
 
-Dense node payload union used by ordered append helpers.
+Initializer tensor names for one single-step recurrent layer.
 
-### PerNeuronLayerParams
+### RecurrentInitializerValues
 
-Parameters for per-neuron layer emission.
+Collected initializer vectors for one single-step recurrent layer.
 
-### PerNeuronLayerContext
+### RecurrentLayerEmissionContext
 
-Per-neuron layer context alias.
+Derived execution context for single-step recurrent layer emission.
 
-### PerNeuronSubgraphContext
+### RecurrentLayerEmissionParams
 
-Per-neuron subgraph emission context.
+Parameters for single-step recurrent layer emission.
 
-### PerNeuronNodeContext
+### RecurrentRowCollectionContext
 
-Per-neuron normalized node context.
+Context for collecting one recurrent matrix row.
 
-### PerNeuronTensorNames
+### SharedActivationNodeBuildParams
 
-Per-neuron initializer tensor names.
+Shared parameters for constructing an activation node payload.
 
-### PerNeuronGraphNames
+### SharedGemmNodeBuildParams
 
-Per-neuron graph tensor names.
+Shared parameters for constructing a Gemm node payload.
 
-### PerNeuronConcatNodePayload
+### SpecMetadataAppendContext
 
-Per-neuron concat node payload.
+Append-a-spec metadata context for JSON-array metadata keys.
+
+### WeightToleranceComparisonContext
+
+Context for comparing two scalar weights with numeric tolerance.
 
 ## architecture/network/onnx/export/network.onnx.export-flow.utils.ts
 
@@ -443,35 +443,39 @@ Returns: ONNX model.
 
 ## architecture/network/onnx/export/network.onnx.export-setup.utils.ts
 
-### createGraphDimensions
+### appendRecurrentGraphInput
 
 ```ts
-createGraphDimensions(
-  context: OnnxGraphDimensionBuildContext,
-): OnnxGraphDimensions
+appendRecurrentGraphInput(
+  model: OnnxModel,
+  traversalContext: OnnxRecurrentLayerTraversalContext,
+): void
 ```
 
-Build tensor dimensions for model input and output, optionally with symbolic batch dimension.
+Append one recurrent previous-state graph input for a hidden layer.
 
 Parameters:
-- `context` - Dimension construction context.
+- `model` - Target ONNX model.
+- `traversalContext` - Hidden layer traversal context.
 
-Returns: Input and output dimension arrays for ONNX value info.
+Returns: Nothing.
 
-### createBaseModel
+### appendRecurrentLayerIndex
 
 ```ts
-createBaseModel(
-  context: OnnxBaseModelBuildContext,
-): OnnxModel
+appendRecurrentLayerIndex(
+  recurrentLayerIndices: number[],
+  traversalContext: OnnxRecurrentLayerTraversalContext,
+): void
 ```
 
-Create the base ONNX model shell with graph input/output declarations.
+Append one recurrent layer index to the collected index list.
 
 Parameters:
-- `context` - Base model build context.
+- `recurrentLayerIndices` - Collected recurrent layer indices.
+- `traversalContext` - Hidden layer traversal context.
 
-Returns: Initialized ONNX model with empty initializer/node lists.
+Returns: Nothing.
 
 ### applyModelMetadata
 
@@ -503,22 +507,35 @@ Parameters:
 
 Returns: Export-layer indices with recurrent self-connections.
 
-### createTensorDimensions
+### createBaseModel
 
 ```ts
-createTensorDimensions(
-  width: number,
-  batchDimension: boolean,
-): OnnxDimension[]
+createBaseModel(
+  context: OnnxBaseModelBuildContext,
+): OnnxModel
+```
+
+Create the base ONNX model shell with graph input/output declarations.
+
+Parameters:
+- `context` - Base model build context.
+
+Returns: Initialized ONNX model with empty initializer/node lists.
+
+### createGraphDimensions
+
+```ts
+createGraphDimensions(
+  context: OnnxGraphDimensionBuildContext,
+): OnnxGraphDimensions
 ```
 
-Build one tensor shape dimension payload for dense vectors.
+Build tensor dimensions for model input and output, optionally with symbolic batch dimension.
 
 Parameters:
-- `width` - Vector width.
-- `batchDimension` - Whether symbolic batch dimension is enabled.
+- `context` - Dimension construction context.
 
-Returns: ONNX dimensions for the vector payload.
+Returns: Input and output dimension arrays for ONNX value info.
 
 ### createGraphValueInfo
 
@@ -537,20 +554,20 @@ Parameters:
 
 Returns: ONNX value info payload.
 
-### isRecurrentCollectionEnabled
+### createHiddenLayerIndices
 
 ```ts
-isRecurrentCollectionEnabled(
-  context: OnnxRecurrentCollectionContext,
-): boolean
+createHiddenLayerIndices(
+  totalLayerCount: number,
+): number[]
 ```
 
-Determine whether recurrent layer collection should execute.
+Build hidden layer indices excluding input and output layers.
 
 Parameters:
-- `context` - Recurrent collection context.
+- `totalLayerCount` - Total number of network layers.
 
-Returns: True when recurrent collection is enabled.
+Returns: Hidden layer indices.
 
 ### createHiddenLayerTraversalContexts
 
@@ -567,69 +584,52 @@ Parameters:
 
 Returns: Hidden layer traversal contexts.
 
-### createHiddenLayerIndices
-
-```ts
-createHiddenLayerIndices(
-  totalLayerCount: number,
-): number[]
-```
-
-Build hidden layer indices excluding input and output layers.
-
-Parameters:
-- `totalLayerCount` - Total number of network layers.
-
-Returns: Hidden layer indices.
-
-### processHiddenLayerRecurrence
+### createRecurrentInputValueInfo
 
 ```ts
-processHiddenLayerRecurrence(
-  context: OnnxRecurrentLayerProcessingContext,
-): void
+createRecurrentInputValueInfo(
+  context: OnnxRecurrentInputValueInfoContext,
+): OnnxValueInfo
 ```
 
-Process one hidden layer for recurrent self-connections.
+Build one recurrent previous-state graph input payload.
 
 Parameters:
-- `context` - Hidden layer recurrent processing context.
+- `context` - Recurrent input value-info context.
 
-Returns: Nothing.
+Returns: ONNX value info payload for recurrent state input.
 
-### appendRecurrentLayerIndex
+### createRecurrentInputValueInfoContext
 
 ```ts
-appendRecurrentLayerIndex(
-  recurrentLayerIndices: number[],
+createRecurrentInputValueInfoContext(
   traversalContext: OnnxRecurrentLayerTraversalContext,
-): void
+): OnnxRecurrentInputValueInfoContext
 ```
 
-Append one recurrent layer index to the collected index list.
+Build recurrent input context for one hidden recurrent layer.
 
 Parameters:
-- `recurrentLayerIndices` - Collected recurrent layer indices.
 - `traversalContext` - Hidden layer traversal context.
 
-Returns: Nothing.
+Returns: Recurrent input value-info context.
 
-### appendRecurrentGraphInput
+### createTensorDimensions
 
 ```ts
-appendRecurrentGraphInput(
-  model: OnnxModel,
-  traversalContext: OnnxRecurrentLayerTraversalContext,
-): void
+createTensorDimensions(
+  width: number,
+  batchDimension: boolean,
+): OnnxDimension[]
 ```
 
-Append one recurrent previous-state graph input for a hidden layer.
+Build one tensor shape dimension payload for dense vectors.
 
 Parameters:
-- `model` - Target ONNX model.
-- `traversalContext` - Hidden layer traversal context.
+- `width` - Vector width.
+- `batchDimension` - Whether symbolic batch dimension is enabled.
 
-Returns: Nothing.
+Returns: ONNX dimensions for the vector payload.
 
 ### hasLayerSelfRecurrence
 
@@ -646,200 +646,197 @@ Parameters:
 
 Returns: True when a node has a self-connection.
 
-### createRecurrentInputValueInfoContext
+### isRecurrentCollectionEnabled
 
 ```ts
-createRecurrentInputValueInfoContext(
-  traversalContext: OnnxRecurrentLayerTraversalContext,
-): OnnxRecurrentInputValueInfoContext
+isRecurrentCollectionEnabled(
+  context: OnnxRecurrentCollectionContext,
+): boolean
 ```
 
-Build recurrent input context for one hidden recurrent layer.
+Determine whether recurrent layer collection should execute.
 
 Parameters:
-- `traversalContext` - Hidden layer traversal context.
+- `context` - Recurrent collection context.
 
-Returns: Recurrent input value-info context.
+Returns: True when recurrent collection is enabled.
 
-### createRecurrentInputValueInfo
+### processHiddenLayerRecurrence
 
 ```ts
-createRecurrentInputValueInfo(
-  context: OnnxRecurrentInputValueInfoContext,
-): OnnxValueInfo
+processHiddenLayerRecurrence(
+  context: OnnxRecurrentLayerProcessingContext,
+): void
 ```
 
-Build one recurrent previous-state graph input payload.
+Process one hidden layer for recurrent self-connections.
 
 Parameters:
-- `context` - Recurrent input value-info context.
+- `context` - Hidden layer recurrent processing context.
 
-Returns: ONNX value info payload for recurrent state input.
+Returns: Nothing.
 
 ## architecture/network/onnx/export/network.onnx.export-postprocess.utils.ts
 
-### emitFusedRecurrentHeuristics
+### appendConvLayerValidationResult
 
 ```ts
-emitFusedRecurrentHeuristics(
-  model: OnnxModel,
-  layers: default[][],
-  allowRecurrent: boolean | undefined,
-  previousOutputName: string,
+appendConvLayerValidationResult(
+  result: ConvSharingValidationResult,
+  layerIndex: number,
+  isConsistent: boolean,
 ): void
 ```
 
-Emit heuristic fused recurrent operators (LSTM/GRU) when recurrent export is enabled.
-
-Parameters:
-- `model` - Target ONNX model.
-- `layers` - Layered network nodes.
-- `allowRecurrent` - Whether recurrent export is enabled.
-- `previousOutputName` - Current graph output name (kept for backward-compatible emission semantics).
-
-Returns: Nothing.
+Append one Conv-layer validation outcome and optional warning.
 
-### finalizeExportMetadata
+### appendConvSharingMetadata
 
 ```ts
-finalizeExportMetadata(
+appendConvSharingMetadata(
   model: OnnxModel,
-  layers: default[][],
-  options: OnnxExportOptions,
-  includeMetadata: boolean,
-  hiddenSizesMetadata: number[],
-  recurrentLayerIndices: number[],
+  result: ConvSharingValidationResult,
 ): void
 ```
 
-Finalize export metadata and optional conv-sharing validation.
+Append Conv-sharing validation metadata arrays.
 
-Parameters:
-- `model` - Target ONNX model.
-- `layers` - Layered network nodes.
-- `options` - Export options.
-- `includeMetadata` - Whether metadata emission is enabled.
-- `hiddenSizesMetadata` - Hidden-layer sizes collected during emission.
-- `recurrentLayerIndices` - Recurrent layer indices.
+### appendFusedRecurrentInitializers
 
-Returns: Nothing.
+```ts
+appendFusedRecurrentInitializers(
+  model: OnnxModel,
+  initializerNames: FusedRecurrentInitializerNames,
+  parameters: RecurrentGateParameterCollectionResult,
+  gateCount: number,
+  unitSize: number,
+  previousSize: number,
+): void
+```
 
-### tryEmitFusedLstm
+Append fused recurrent initializer tensors to the ONNX graph.
+
+### appendFusedRecurrentNode
 
 ```ts
-tryEmitFusedLstm(
-  context: HiddenLayerHeuristicContext,
+appendFusedRecurrentNode(
+  graph: OnnxGraph,
+  operatorType: "LSTM" | "GRU",
+  previousOutputName: string,
+  initializerNames: FusedRecurrentInitializerNames,
+  graphNames: FusedRecurrentGraphNames,
+  unitSize: number,
 ): void
 ```
 
-Try emitting heuristic fused LSTM node and metadata.
+Append fused recurrent operator node to the ONNX graph.
 
-### buildFusedLstmExecutionContext
+### appendIndexMetadata
 
 ```ts
-buildFusedLstmExecutionContext(
-  context: LstmEmissionContext,
-): FusedRecurrentEmissionExecutionContext
+appendIndexMetadata(
+  model: OnnxModel,
+  key: string,
+  layerIndex: number,
+): void
 ```
 
-Build shared fused-recurrent execution context for LSTM.
+Append a unique layer index to metadata array key.
 
-### tryEmitFusedGru
+### appendMetadataProperty
 
 ```ts
-tryEmitFusedGru(
-  context: HiddenLayerHeuristicContext,
+appendMetadataProperty(
+  model: OnnxModel,
+  metadataProperty: OnnxMetadataProperty,
 ): void
 ```
 
-Try emitting heuristic fused GRU node and metadata.
+Append metadata property to model metadata_props list.
 
-### buildFusedGruExecutionContext
+### appendRecurrentSingleStepMetadata
 
 ```ts
-buildFusedGruExecutionContext(
-  context: GruEmissionContext,
-): FusedRecurrentEmissionExecutionContext
+appendRecurrentSingleStepMetadata(
+  model: OnnxModel,
+  recurrentLayerIndices: number[],
+): void
 ```
 
-Build shared fused-recurrent execution context for GRU.
+Append recurrent single-step metadata when recurrent layers exist.
 
-### emitFusedRecurrentLayer
+### areWeightsWithinTolerance
 
 ```ts
-emitFusedRecurrentLayer(
-  context: FusedRecurrentEmissionExecutionContext,
-): void
+areWeightsWithinTolerance(
+  context: WeightToleranceComparisonContext,
+): boolean
 ```
 
-Emit shared fused recurrent payload (initializers, node, metadata).
+Compare two scalar weights using configured tolerance.
 
-### appendIndexMetadata
+### asNodeInternals
 
 ```ts
-appendIndexMetadata(
-  model: OnnxModel,
-  key: string,
-  layerIndex: number,
-): void
+asNodeInternals(
+  node: default,
+): NodeInternals
 ```
 
-Append a unique layer index to metadata array key.
+Resolve runtime node internals in one typed helper.
 
-### findMetadataPropertyIndex
+### buildFusedGruExecutionContext
 
 ```ts
-findMetadataPropertyIndex(
-  metadataProperties: OnnxMetadataProperty[],
-  key: string,
-): number
+buildFusedGruExecutionContext(
+  context: GruEmissionContext,
+): FusedRecurrentEmissionExecutionContext
 ```
 
-Find metadata property index by key.
+Build shared fused-recurrent execution context for GRU.
 
-### upsertLayerIndexMetadataValue
+### buildFusedLstmExecutionContext
 
 ```ts
-upsertLayerIndexMetadataValue(
-  metadataProperties: OnnxMetadataProperty[],
-  metadataIndex: number,
-  layerIndex: number,
-): void
+buildFusedLstmExecutionContext(
+  context: LstmEmissionContext,
+): FusedRecurrentEmissionExecutionContext
 ```
 
-Upsert one layer index into metadata array-like JSON value.
+Build shared fused-recurrent execution context for LSTM.
 
-### parseMetadataLayerIndices
+### buildFusedRecurrentGraphNames
 
 ```ts
-parseMetadataLayerIndices(
-  metadataValue: string,
-): number[]
+buildFusedRecurrentGraphNames(
+  nodePrefix: string,
+  outputSuffix: string,
+  layerIndex: number,
+): FusedRecurrentGraphNames
 ```
 
-Parse metadata JSON value into a numeric layer-index array.
+Build fused recurrent graph names for node and output.
 
-### buildRecurrentHeuristicEmissionContext
+### buildFusedRecurrentInitializerNames
 
 ```ts
-buildRecurrentHeuristicEmissionContext(
-  model: OnnxModel,
-  layers: default[][],
-  previousOutputName: string,
-): RecurrentHeuristicEmissionContext
+buildFusedRecurrentInitializerNames(
+  operatorType: "LSTM" | "GRU",
+  layerIndex: number,
+): FusedRecurrentInitializerNames
 ```
 
-Build reusable context for recurrent heuristic traversal.
+Build fused recurrent initializer names for the current layer.
 
-### collectHiddenLayerIndices
+### buildGruEmissionContext
 
 ```ts
-collectHiddenLayerIndices(
-  layers: default[][],
-): number[]
+buildGruEmissionContext(
+  context: HiddenLayerHeuristicContext,
+): GruEmissionContext
 ```
 
-Collect hidden-layer indices for recurrent traversal.
+Build GRU emission context from one hidden-layer traversal record.
 
 ### buildHiddenLayerHeuristicContext
 
@@ -852,85 +849,88 @@ buildHiddenLayerHeuristicContext(
 
 Build one hidden-layer traversal context.
 
-### emitFallbackRecurrentPatternMetadata
+### buildLstmEmissionContext
 
 ```ts
-emitFallbackRecurrentPatternMetadata(
+buildLstmEmissionContext(
   context: HiddenLayerHeuristicContext,
-): void
+): LstmEmissionContext
 ```
 
-Emit fallback metadata for recurrent-size ambiguity.
+Build LSTM emission context from one hidden-layer traversal record.
 
-### isFallbackRecurrentPatternSize
+### buildMetadataProperty
 
 ```ts
-isFallbackRecurrentPatternSize(
-  currentSize: number,
-): boolean
+buildMetadataProperty(
+  key: string,
+  value: unknown,
+): OnnxMetadataProperty
 ```
 
-Check whether hidden size should emit recurrent fallback metadata.
+Build a metadata key/value property with JSON string serialization.
 
-### isEligibleForLstmHeuristic
+### buildRecurrentHeuristicEmissionContext
 
 ```ts
-isEligibleForLstmHeuristic(
-  currentSize: number,
-): boolean
+buildRecurrentHeuristicEmissionContext(
+  model: OnnxModel,
+  layers: default[][],
+  previousOutputName: string,
+): RecurrentHeuristicEmissionContext
 ```
 
-Check LSTM heuristic eligibility by size and gate divisibility.
+Build reusable context for recurrent heuristic traversal.
 
-### buildLstmEmissionContext
+### collectConvKernelCoordinates
 
 ```ts
-buildLstmEmissionContext(
-  context: HiddenLayerHeuristicContext,
-): LstmEmissionContext
+collectConvKernelCoordinates(
+  convSpec: Conv2DMapping,
+): OnnxConvKernelCoordinate[]
 ```
 
-Build LSTM emission context from one hidden-layer traversal record.
+Collect kernel coordinates for one Conv kernel traversal.
 
-### collectLstmGateNodeGroups
+### collectConvOutputCoordinates
 
 ```ts
-collectLstmGateNodeGroups(
-  context: LstmEmissionContext,
-): default[][]
+collectConvOutputCoordinates(
+  convSpec: Conv2DMapping,
+): ConvOutputCoordinate[]
 ```
 
-Collect LSTM gate node groups in canonical export order.
+Collect output coordinates for full Conv traversal.
 
-### isEligibleForGruHeuristic
+### collectGruGateNodeGroups
 
 ```ts
-isEligibleForGruHeuristic(
-  currentSize: number,
-): boolean
+collectGruGateNodeGroups(
+  context: GruEmissionContext,
+): default[][]
 ```
 
-Check GRU heuristic eligibility by size and gate divisibility.
+Collect GRU gate node groups in canonical export order.
 
-### buildGruEmissionContext
+### collectHiddenLayerIndices
 
 ```ts
-buildGruEmissionContext(
-  context: HiddenLayerHeuristicContext,
-): GruEmissionContext
+collectHiddenLayerIndices(
+  layers: default[][],
+): number[]
 ```
 
-Build GRU emission context from one hidden-layer traversal record.
+Collect hidden-layer indices for recurrent traversal.
 
-### collectGruGateNodeGroups
+### collectLstmGateNodeGroups
 
 ```ts
-collectGruGateNodeGroups(
-  context: GruEmissionContext,
+collectLstmGateNodeGroups(
+  context: LstmEmissionContext,
 ): default[][]
 ```
 
-Collect GRU gate node groups in canonical export order.
+Collect LSTM gate node groups in canonical export order.
 
 ### collectRecurrentGateBlockParameters
 
@@ -952,142 +952,145 @@ collectRecurrentGateRow(
 
 Collect one recurrent gate row payload (inputs, recurrent slice, and bias).
 
-### resolveRecurrentRowWeight
-
-```ts
-resolveRecurrentRowWeight(
-  context: RecurrentGateRowCollectionContext,
-  columnIndex: number,
-): number
-```
-
-Resolve one recurrent row value at the requested column.
-
-### foldRecurrentGateRows
+### collectRepresentativeKernelForChannel
 
 ```ts
-foldRecurrentGateRows(
-  gateRows: RecurrentGateRow[],
-): RecurrentGateParameterCollectionResult
+collectRepresentativeKernelForChannel(
+  context: ConvRepresentativeKernelContext,
+): number[]
 ```
 
-Fold recurrent gate rows into flattened ONNX initializer vectors.
+Collect one representative kernel by reading the first output position for a channel.
 
-### foldRecurrentGateBlocks
+### collectRepresentativeKernels
 
 ```ts
-foldRecurrentGateBlocks(
-  gateParameterBlocks: RecurrentGateParameterCollectionResult[],
-): RecurrentGateParameterCollectionResult
+collectRepresentativeKernels(
+  context: ConvLayerPairContext,
+): number[][]
 ```
 
-Fold gate blocks into a single fused parameter payload.
+Collect representative kernels for each output channel.
 
-### buildFusedRecurrentInitializerNames
+### collectRepresentativeKernelWeight
 
 ```ts
-buildFusedRecurrentInitializerNames(
-  operatorType: "LSTM" | "GRU",
-  layerIndex: number,
-): FusedRecurrentInitializerNames
+collectRepresentativeKernelWeight(
+  convSpec: Conv2DMapping,
+  previousLayerNodes: default[],
+  representativeInternal: NodeInternals,
+  kernelCoordinate: OnnxConvKernelCoordinate,
+): number
 ```
 
-Build fused recurrent initializer names for the current layer.
+Collect representative kernel value using top-left receptive field indexing.
 
-### buildFusedRecurrentGraphNames
+### emitFallbackRecurrentPatternMetadata
 
 ```ts
-buildFusedRecurrentGraphNames(
-  nodePrefix: string,
-  outputSuffix: string,
-  layerIndex: number,
-): FusedRecurrentGraphNames
+emitFallbackRecurrentPatternMetadata(
+  context: HiddenLayerHeuristicContext,
+): void
 ```
 
-Build fused recurrent graph names for node and output.
+Emit fallback metadata for recurrent-size ambiguity.
 
-### appendFusedRecurrentInitializers
+### emitFusedRecurrentHeuristics
 
 ```ts
-appendFusedRecurrentInitializers(
+emitFusedRecurrentHeuristics(
   model: OnnxModel,
-  initializerNames: FusedRecurrentInitializerNames,
-  parameters: RecurrentGateParameterCollectionResult,
-  gateCount: number,
-  unitSize: number,
-  previousSize: number,
+  layers: default[][],
+  allowRecurrent: boolean | undefined,
+  previousOutputName: string,
 ): void
 ```
 
-Append fused recurrent initializer tensors to the ONNX graph.
+Emit heuristic fused recurrent operators (LSTM/GRU) when recurrent export is enabled.
 
-### appendFusedRecurrentNode
+Parameters:
+- `model` - Target ONNX model.
+- `layers` - Layered network nodes.
+- `allowRecurrent` - Whether recurrent export is enabled.
+- `previousOutputName` - Current graph output name (kept for backward-compatible emission semantics).
+
+Returns: Nothing.
+
+### emitFusedRecurrentLayer
 
 ```ts
-appendFusedRecurrentNode(
-  graph: OnnxGraph,
-  operatorType: "LSTM" | "GRU",
-  previousOutputName: string,
-  initializerNames: FusedRecurrentInitializerNames,
-  graphNames: FusedRecurrentGraphNames,
-  unitSize: number,
+emitFusedRecurrentLayer(
+  context: FusedRecurrentEmissionExecutionContext,
 ): void
 ```
 
-Append fused recurrent operator node to the ONNX graph.
+Emit shared fused recurrent payload (initializers, node, metadata).
 
-### resolveGruPreviousOutputName
+### ensureMetadataProps
 
 ```ts
-resolveGruPreviousOutputName(
-  layerIndex: number,
-): string
+ensureMetadataProps(
+  model: OnnxModel,
+): OnnxMetadataProperty[]
 ```
 
-Resolve previous output naming semantics for GRU heuristic emission.
+Ensure metadata_props array exists and return it.
 
-### appendRecurrentSingleStepMetadata
+### finalizeExportMetadata
 
 ```ts
-appendRecurrentSingleStepMetadata(
+finalizeExportMetadata(
   model: OnnxModel,
+  layers: default[][],
+  options: OnnxExportOptions,
+  includeMetadata: boolean,
+  hiddenSizesMetadata: number[],
   recurrentLayerIndices: number[],
 ): void
 ```
 
-Append recurrent single-step metadata when recurrent layers exist.
+Finalize export metadata and optional conv-sharing validation.
 
-### shouldValidateConvSharing
+Parameters:
+- `model` - Target ONNX model.
+- `layers` - Layered network nodes.
+- `options` - Export options.
+- `includeMetadata` - Whether metadata emission is enabled.
+- `hiddenSizesMetadata` - Hidden-layer sizes collected during emission.
+- `recurrentLayerIndices` - Recurrent layer indices.
+
+Returns: Nothing.
+
+### findMetadataPropertyIndex
 
 ```ts
-shouldValidateConvSharing(
-  options: OnnxExportOptions,
-): boolean
+findMetadataPropertyIndex(
+  metadataProperties: OnnxMetadataProperty[],
+  key: string,
+): number
 ```
 
-Determine whether Conv2D sharing validation is enabled and configured.
-
-### validateConvSharingAcrossMappings
+Find metadata property index by key.
 
-```ts
-validateConvSharingAcrossMappings(
-  context: ConvSharingValidationContext,
-): ConvSharingValidationResult
+### foldRecurrentGateBlocks
+
+```ts
+foldRecurrentGateBlocks(
+  gateParameterBlocks: RecurrentGateParameterCollectionResult[],
+): RecurrentGateParameterCollectionResult
 ```
 
-Validate Conv2D sharing across all declared Conv mappings.
+Fold gate blocks into a single fused parameter payload.
 
-### resolveConvLayerPairContext
+### foldRecurrentGateRows
 
 ```ts
-resolveConvLayerPairContext(
-  layers: default[][],
-  layerIndex: number,
-  convSpec: Conv2DMapping,
-): ConvLayerPairContext | undefined
+foldRecurrentGateRows(
+  gateRows: RecurrentGateRow[],
+): RecurrentGateParameterCollectionResult
 ```
 
-Resolve one Conv mapping layer pair or return undefined for invalid layout.
+Fold recurrent gate rows into flattened ONNX initializer vectors.
 
 ### isConvLayerPairConsistent
 
@@ -1099,68 +1102,57 @@ isConvLayerPairConsistent(
 
 Validate one Conv layer pair against representative kernel sharing.
 
-### appendConvLayerValidationResult
-
-```ts
-appendConvLayerValidationResult(
-  result: ConvSharingValidationResult,
-  layerIndex: number,
-  isConsistent: boolean,
-): void
-```
-
-Append one Conv-layer validation outcome and optional warning.
-
-### appendConvSharingMetadata
+### isEligibleForGruHeuristic
 
 ```ts
-appendConvSharingMetadata(
-  model: OnnxModel,
-  result: ConvSharingValidationResult,
-): void
+isEligibleForGruHeuristic(
+  currentSize: number,
+): boolean
 ```
 
-Append Conv-sharing validation metadata arrays.
+Check GRU heuristic eligibility by size and gate divisibility.
 
-### collectRepresentativeKernels
+### isEligibleForLstmHeuristic
 
 ```ts
-collectRepresentativeKernels(
-  context: ConvLayerPairContext,
-): number[][]
+isEligibleForLstmHeuristic(
+  currentSize: number,
+): boolean
 ```
 
-Collect representative kernels for each output channel.
+Check LSTM heuristic eligibility by size and gate divisibility.
 
-### collectRepresentativeKernelForChannel
+### isFallbackRecurrentPatternSize
 
 ```ts
-collectRepresentativeKernelForChannel(
-  context: ConvRepresentativeKernelContext,
-): number[]
+isFallbackRecurrentPatternSize(
+  currentSize: number,
+): boolean
 ```
 
-Collect one representative kernel by reading the first output position for a channel.
+Check whether hidden size should emit recurrent fallback metadata.
 
-### collectConvOutputCoordinates
+### isInputPositionInsideBounds
 
 ```ts
-collectConvOutputCoordinates(
+isInputPositionInsideBounds(
   convSpec: Conv2DMapping,
-): ConvOutputCoordinate[]
+  inputRow: number,
+  inputColumn: number,
+): boolean
 ```
 
-Collect output coordinates for full Conv traversal.
+Check whether input row/column falls inside Conv input bounds.
 
-### collectConvKernelCoordinates
+### isKernelCoordinateConsistent
 
 ```ts
-collectConvKernelCoordinates(
-  convSpec: Conv2DMapping,
-): OnnxConvKernelCoordinate[]
+isKernelCoordinateConsistent(
+  context: ConvKernelConsistencyContext,
+): boolean
 ```
 
-Collect kernel coordinates for one Conv kernel traversal.
+Validate one kernel coordinate against its representative channel value.
 
 ### isOutputCoordinateConsistent
 
@@ -1175,106 +1167,80 @@ isOutputCoordinateConsistent(
 
 Validate one output coordinate against channel representative kernel weights.
 
-### resolveNeuronInternalAtOutputCoordinate
-
-```ts
-resolveNeuronInternalAtOutputCoordinate(
-  context: ConvLayerPairContext,
-  outputCoordinate: ConvOutputCoordinate,
-): NodeInternals | undefined
-```
-
-Resolve runtime internals for output coordinate neuron, if present.
-
-### isKernelCoordinateConsistent
-
-```ts
-isKernelCoordinateConsistent(
-  context: ConvKernelConsistencyContext,
-): boolean
-```
-
-Validate one kernel coordinate against its representative channel value.
-
-### resolveInputPosition
+### parseMetadataLayerIndices
 
 ```ts
-resolveInputPosition(
-  context: ConvKernelConsistencyContext,
-): { inputRow: number; inputColumn: number; }
+parseMetadataLayerIndices(
+  metadataValue: string,
+): number[]
 ```
 
-Resolve input row/column projected by output and kernel coordinates.
+Parse metadata JSON value into a numeric layer-index array.
 
-### isInputPositionInsideBounds
+### resolveConvLayerPairContext
 
 ```ts
-isInputPositionInsideBounds(
+resolveConvLayerPairContext(
+  layers: default[][],
+  layerIndex: number,
   convSpec: Conv2DMapping,
-  inputRow: number,
-  inputColumn: number,
-): boolean
+): ConvLayerPairContext | undefined
 ```
 
-Check whether input row/column falls inside Conv input bounds.
+Resolve one Conv mapping layer pair or return undefined for invalid layout.
 
-### resolveSourceNodeAtInputPosition
+### resolveGruPreviousOutputName
 
 ```ts
-resolveSourceNodeAtInputPosition(
-  convSpec: Conv2DMapping,
-  previousLayerNodes: default[],
-  inChannelIndex: number,
-  inputRow: number,
-  inputColumn: number,
-): default | undefined
+resolveGruPreviousOutputName(
+  layerIndex: number,
+): string
 ```
 
-Resolve source node by Conv input position coordinates.
+Resolve previous output naming semantics for GRU heuristic emission.
 
-### collectRepresentativeKernelWeight
+### resolveIncomingWeight
 
 ```ts
-collectRepresentativeKernelWeight(
-  convSpec: Conv2DMapping,
-  previousLayerNodes: default[],
-  representativeInternal: NodeInternals,
-  kernelCoordinate: OnnxConvKernelCoordinate,
+resolveIncomingWeight(
+  targetNodeInternal: NodeInternals,
+  sourceNode: default,
 ): number
 ```
 
-Collect representative kernel value using top-left receptive field indexing.
+Resolve incoming connection weight from a specific source node.
 
-### areWeightsWithinTolerance
+### resolveInputPosition
 
 ```ts
-areWeightsWithinTolerance(
-  context: WeightToleranceComparisonContext,
-): boolean
+resolveInputPosition(
+  context: ConvKernelConsistencyContext,
+): { inputRow: number; inputColumn: number; }
 ```
 
-Compare two scalar weights using configured tolerance.
+Resolve input row/column projected by output and kernel coordinates.
 
-### asNodeInternals
+### resolveNeuronInternalAtOutputCoordinate
 
 ```ts
-asNodeInternals(
-  node: default,
-): NodeInternals
+resolveNeuronInternalAtOutputCoordinate(
+  context: ConvLayerPairContext,
+  outputCoordinate: ConvOutputCoordinate,
+): NodeInternals | undefined
 ```
 
-Resolve runtime node internals in one typed helper.
+Resolve runtime internals for output coordinate neuron, if present.
 
-### resolveIncomingWeight
+### resolveRecurrentRowWeight
 
 ```ts
-resolveIncomingWeight(
-  targetNodeInternal: NodeInternals,
-  sourceNode: default,
+resolveRecurrentRowWeight(
+  context: RecurrentGateRowCollectionContext,
+  columnIndex: number,
 ): number
 ```
 
-Resolve incoming connection weight from a specific source node.
+Resolve one recurrent row value at the requested column.
 
 ### resolveSelfConnectionWeight
 
@@ -1286,71 +1252,73 @@ resolveSelfConnectionWeight(
 
 Resolve self-connection weight for diagonal recurrent matrix entries.
 
-### buildMetadataProperty
+### resolveSourceNodeAtInputPosition
 
 ```ts
-buildMetadataProperty(
-  key: string,
-  value: unknown,
-): OnnxMetadataProperty
+resolveSourceNodeAtInputPosition(
+  convSpec: Conv2DMapping,
+  previousLayerNodes: default[],
+  inChannelIndex: number,
+  inputRow: number,
+  inputColumn: number,
+): default | undefined
 ```
 
-Build a metadata key/value property with JSON string serialization.
+Resolve source node by Conv input position coordinates.
 
-### appendMetadataProperty
+### shouldValidateConvSharing
 
 ```ts
-appendMetadataProperty(
-  model: OnnxModel,
-  metadataProperty: OnnxMetadataProperty,
-): void
+shouldValidateConvSharing(
+  options: OnnxExportOptions,
+): boolean
 ```
 
-Append metadata property to model metadata_props list.
+Determine whether Conv2D sharing validation is enabled and configured.
 
-### ensureMetadataProps
+### tryEmitFusedGru
 
 ```ts
-ensureMetadataProps(
-  model: OnnxModel,
-): OnnxMetadataProperty[]
+tryEmitFusedGru(
+  context: HiddenLayerHeuristicContext,
+): void
 ```
 
-Ensure metadata_props array exists and return it.
-
-## architecture/network/onnx/export/network.onnx.export-orchestrators.utils.ts
+Try emitting heuristic fused GRU node and metadata.
 
-### assignExportNodeIndices
+### tryEmitFusedLstm
 
 ```ts
-assignExportNodeIndices(
-  network: default,
+tryEmitFusedLstm(
+  context: HiddenLayerHeuristicContext,
 ): void
 ```
 
-Assign stable index values to nodes for export diagnostics.
+Try emitting heuristic fused LSTM node and metadata.
 
-Parameters:
-- `network` - Source network.
+### upsertLayerIndexMetadataValue
 
-Returns: Nothing.
+```ts
+upsertLayerIndexMetadataValue(
+  metadataProperties: OnnxMetadataProperty[],
+  metadataIndex: number,
+  layerIndex: number,
+): void
+```
 
-### collectLstmPatternStubs
+Upsert one layer index into metadata array-like JSON value.
+
+### validateConvSharingAcrossMappings
 
 ```ts
-collectLstmPatternStubs(
-  layers: default[][],
-  allowRecurrent: boolean | undefined,
-): LstmPatternStub[]
+validateConvSharingAcrossMappings(
+  context: ConvSharingValidationContext,
+): ConvSharingValidationResult
 ```
 
-Collect heuristic LSTM grouping stubs from hidden layers.
-
-Parameters:
-- `layers` - Layered network nodes.
-- `allowRecurrent` - Whether recurrent export heuristics are enabled.
+Validate Conv2D sharing across all declared Conv mappings.
 
-Returns: Candidate LSTM pattern stubs.
+## architecture/network/onnx/export/network.onnx.export-orchestrators.utils.ts
 
 ### appendConvInferenceMetadata
 
@@ -1388,20 +1356,22 @@ Parameters:
 
 Returns: Nothing.
 
-### createExportNodeIndexAssignmentContexts
+### appendMetadataProperties
 
 ```ts
-createExportNodeIndexAssignmentContexts(
-  network: default,
-): ExportNodeIndexAssignmentContext[]
+appendMetadataProperties(
+  model: OnnxModel,
+  metadataProperties: OnnxMetadataProperty[],
+): void
 ```
 
-Create node/index assignment contexts for export diagnostics.
+Append metadata properties in a single, normalized path.
 
 Parameters:
-- `network` - Source network.
+- `model` - Target ONNX model.
+- `metadataProperties` - Metadata properties to append.
 
-Returns: Assignment contexts.
+Returns: Nothing.
 
 ### applyExportNodeIndexAssignments
 
@@ -1433,20 +1403,52 @@ Parameters:
 
 Returns: Nothing.
 
-### safelyCollectLstmPatternStubs
+### assignExportNodeIndices
 
 ```ts
-safelyCollectLstmPatternStubs(
+assignExportNodeIndices(
+  network: default,
+): void
+```
+
+Assign stable index values to nodes for export diagnostics.
+
+Parameters:
+- `network` - Source network.
+
+Returns: Nothing.
+
+### collectInferredConvMetadata
+
+```ts
+collectInferredConvMetadata(
+  context: { layers: default[][]; declaredMappings: Conv2DMapping[] | undefined; },
+): ConvInferenceResult
+```
+
+Collect inferred Conv metadata from hidden-layer traversals.
+
+Parameters:
+- `context` - Conv traversal context.
+
+Returns: Inferred Conv metadata result.
+
+### collectLstmPatternStubs
+
+```ts
+collectLstmPatternStubs(
   layers: default[][],
+  allowRecurrent: boolean | undefined,
 ): LstmPatternStub[]
 ```
 
-Collect LSTM pattern stubs with heuristic error isolation.
+Collect heuristic LSTM grouping stubs from hidden layers.
 
 Parameters:
 - `layers` - Layered network nodes.
+- `allowRecurrent` - Whether recurrent export heuristics are enabled.
 
-Returns: LSTM pattern stubs.
+Returns: Candidate LSTM pattern stubs.
 
 ### collectLstmPatternStubsFromLayers
 
@@ -1463,6 +1465,51 @@ Parameters:
 
 Returns: LSTM pattern stubs.
 
+### createConvInferenceEvaluationContext
+
+```ts
+createConvInferenceEvaluationContext(
+  traversalContext: ConvInferenceTraversalContext,
+): ConvInferenceEvaluationContext
+```
+
+Create width/square-evaluation context for Conv inference.
+
+Parameters:
+- `traversalContext` - Conv traversal context.
+
+Returns: Conv evaluation context.
+
+### createConvTraversalContexts
+
+```ts
+createConvTraversalContexts(
+  context: { layers: default[][]; declaredMappings: Conv2DMapping[] | undefined; },
+): ConvInferenceTraversalContext[]
+```
+
+Create Conv traversal contexts for hidden layers.
+
+Parameters:
+- `context` - Conv traversal source context.
+
+Returns: Conv traversal contexts.
+
+### createExportNodeIndexAssignmentContexts
+
+```ts
+createExportNodeIndexAssignmentContexts(
+  network: default,
+): ExportNodeIndexAssignmentContext[]
+```
+
+Create node/index assignment contexts for export diagnostics.
+
+Parameters:
+- `network` - Source network.
+
+Returns: Assignment contexts.
+
 ### createHiddenLayerTraversalContexts
 
 ```ts
@@ -1493,35 +1540,20 @@ Parameters:
 
 Returns: LSTM candidate context.
 
-### isValidLstmCandidateContext
+### hasInferredConvMetadata
 
 ```ts
-isValidLstmCandidateContext(
-  candidateContext: LstmCandidateContext,
+hasInferredConvMetadata(
+  inferenceResult: ConvInferenceResult,
 ): boolean
 ```
 
-Determine whether a candidate context satisfies heuristic LSTM conditions.
-
-Parameters:
-- `candidateContext` - Candidate context.
-
-Returns: True when the candidate is a valid LSTM stub.
-
-### mapLstmCandidateToStub
-
-```ts
-mapLstmCandidateToStub(
-  candidateContext: LstmCandidateContext,
-): LstmPatternStub
-```
-
-Map a valid candidate context to metadata stub.
+Check whether inferred Conv metadata exists.
 
 Parameters:
-- `candidateContext` - Valid candidate context.
+- `inferenceResult` - Inferred Conv result.
 
-Returns: LSTM pattern stub.
+Returns: True when inferred metadata exists.
 
 ### hasRequiredSelfConnectionCount
 
@@ -1538,78 +1570,78 @@ Parameters:
 
 Returns: True when self-connection count matches requirement.
 
-### collectInferredConvMetadata
+### isDeclaredConvLayer
 
 ```ts
-collectInferredConvMetadata(
-  context: { layers: default[][]; declaredMappings: Conv2DMapping[] | undefined; },
-): ConvInferenceResult
+isDeclaredConvLayer(
+  traversalContext: ConvInferenceTraversalContext,
+): boolean
 ```
 
-Collect inferred Conv metadata from hidden-layer traversals.
+Check whether a traversal layer already has declared Conv mapping.
 
 Parameters:
-- `context` - Conv traversal context.
+- `traversalContext` - Conv traversal context.
 
-Returns: Inferred Conv metadata result.
+Returns: True when mapping is already declared.
 
-### createConvTraversalContexts
+### isInferredConvSpec
 
 ```ts
-createConvTraversalContexts(
-  context: { layers: default[][]; declaredMappings: Conv2DMapping[] | undefined; },
-): ConvInferenceTraversalContext[]
+isInferredConvSpec(
+  specification: (Conv2DMapping & { note?: string | undefined; }) | undefined,
+): boolean
 ```
 
-Create Conv traversal contexts for hidden layers.
+Type guard for inferred Conv specifications.
 
 Parameters:
-- `context` - Conv traversal source context.
+- `specification` - Conv specification candidate.
 
-Returns: Conv traversal contexts.
+Returns: True when specification is defined.
 
-### resolveConvInferenceForLayer
+### isValidLstmCandidateContext
 
 ```ts
-resolveConvInferenceForLayer(
-  traversalContext: ConvInferenceTraversalContext,
-): (Conv2DMapping & { note?: string | undefined; }) | undefined
+isValidLstmCandidateContext(
+  candidateContext: LstmCandidateContext,
+): boolean
 ```
 
-Resolve inferred Conv specification for one hidden layer.
+Determine whether a candidate context satisfies heuristic LSTM conditions.
 
 Parameters:
-- `traversalContext` - Conv traversal context.
+- `candidateContext` - Candidate context.
 
-Returns: Inferred Conv specification when matched.
+Returns: True when the candidate is a valid LSTM stub.
 
-### createConvInferenceEvaluationContext
+### mapLstmCandidateToStub
 
 ```ts
-createConvInferenceEvaluationContext(
-  traversalContext: ConvInferenceTraversalContext,
-): ConvInferenceEvaluationContext
+mapLstmCandidateToStub(
+  candidateContext: LstmCandidateContext,
+): LstmPatternStub
 ```
 
-Create width/square-evaluation context for Conv inference.
+Map a valid candidate context to metadata stub.
 
 Parameters:
-- `traversalContext` - Conv traversal context.
+- `candidateContext` - Valid candidate context.
 
-Returns: Conv evaluation context.
+Returns: LSTM pattern stub.
 
-### resolveConvSpecFromKernelCandidates
+### resolveConvInferenceForLayer
 
 ```ts
-resolveConvSpecFromKernelCandidates(
-  evaluationContext: ConvInferenceEvaluationContext,
+resolveConvInferenceForLayer(
+  traversalContext: ConvInferenceTraversalContext,
 ): (Conv2DMapping & { note?: string | undefined; }) | undefined
 ```
 
-Resolve Conv specification using ordered kernel candidates.
+Resolve inferred Conv specification for one hidden layer.
 
 Parameters:
-- `evaluationContext` - Conv evaluation context.
+- `traversalContext` - Conv traversal context.
 
 Returns: Inferred Conv specification when matched.
 
@@ -1628,64 +1660,32 @@ Parameters:
 
 Returns: Inferred Conv specification when matched.
 
-### isDeclaredConvLayer
-
-```ts
-isDeclaredConvLayer(
-  traversalContext: ConvInferenceTraversalContext,
-): boolean
-```
-
-Check whether a traversal layer already has declared Conv mapping.
-
-Parameters:
-- `traversalContext` - Conv traversal context.
-
-Returns: True when mapping is already declared.
-
-### isInferredConvSpec
-
-```ts
-isInferredConvSpec(
-  specification: (Conv2DMapping & { note?: string | undefined; }) | undefined,
-): boolean
-```
-
-Type guard for inferred Conv specifications.
-
-Parameters:
-- `specification` - Conv specification candidate.
-
-Returns: True when specification is defined.
-
-### hasInferredConvMetadata
+### resolveConvSpecFromKernelCandidates
 
 ```ts
-hasInferredConvMetadata(
-  inferenceResult: ConvInferenceResult,
-): boolean
+resolveConvSpecFromKernelCandidates(
+  evaluationContext: ConvInferenceEvaluationContext,
+): (Conv2DMapping & { note?: string | undefined; }) | undefined
 ```
 
-Check whether inferred Conv metadata exists.
+Resolve Conv specification using ordered kernel candidates.
 
 Parameters:
-- `inferenceResult` - Inferred Conv result.
+- `evaluationContext` - Conv evaluation context.
 
-Returns: True when inferred metadata exists.
+Returns: Inferred Conv specification when matched.
 
-### appendMetadataProperties
+### safelyCollectLstmPatternStubs
 
 ```ts
-appendMetadataProperties(
-  model: OnnxModel,
-  metadataProperties: OnnxMetadataProperty[],
-): void
+safelyCollectLstmPatternStubs(
+  layers: default[][],
+): LstmPatternStub[]
 ```
 
-Append metadata properties in a single, normalized path.
+Collect LSTM pattern stubs with heuristic error isolation.
 
 Parameters:
-- `model` - Target ONNX model.
-- `metadataProperties` - Metadata properties to append.
+- `layers` - Layered network nodes.
 
-Returns: Nothing.
+Returns: LSTM pattern stubs.