Module dataflow 

Expression-level dataflow, dtype, and gradient-connectivity analysis.

Builds a serializable graph over ordinary crate-local free functions and inherent methods starting from a chosen entrypoint. Transfer rules live in crate::op_semantics; anything not covered is left GradState::Unknown / AbstractDtype::Unknown rather than guessed.

Structs

DataflowDiagnostic

DtypeConflict

DtypeRisk

A same-dtype op where at least one operand is known and another remains unknown.

EdgeId

ExprEdge

ExprGraph

ExprNode

NodeId

NumericDomainViolation

A partial function applied to a domain that can attain a forbidden endpoint.

ZeroTimesInfinity

Multiply that can evaluate 0 * ±inf (silent NaN) rather than a loud -inf loss.

Enums

EdgeKind

Edge role in the expression graph.

GradState

Gradient / trainability state attached to an expression node.

NodeKind

What an expression node represents in source.

NumericImpact

How a numeric hazard can interfere with training or inference.

Functions

analyze

Analyze entrypoint against a loaded crate.

analyze_with_candle_version

Analyze with the resolved candle-nn version used to gate version-specific gradient rules.

analyze_with_phase

Analyze an entrypoint for a specific execution phase (train autograd vs inference no-grad).