Struct CompiledGraph 

pub struct CompiledGraph {
    pub nodes: Vec<NodeEntry>,
    pub free_keys: Vec<String>,
    pub node_free_cols: Vec<Vec<(usize, usize)>>,
    pub tied_plan: TiedPlan,
    pub dataset_offsets: Vec<usize>,
}

The compiled representation of a FitGraphSpec, ready for evaluation.

Fields

nodes: Vec<NodeEntry>

Nodes in their original declaration order (used for evaluation).

free_keys: Vec<String>

Keys "node_id.param_name" for all free parameters, sorted by node_id alphabetically, then by model param order within a node.

node_free_cols: Vec<Vec<(usize, usize)>>

Per-node Jacobian column layout, pre-computed during [compile].

node_free_cols[i] lists (local_param_idx, jac_col) for every free parameter on nodes[i]. Used by the solver to avoid string-parsing free_keys on every iteration.

tied_plan: TiedPlan

Dependency-ordered plan for tied (expr_edge) parameters.

Empty when the graph declares no expr_edges. When non-empty, the solver calls TiedPlan::apply on the flat parameter map after updating the free parameters on every iteration, so that each tied target is recomputed from its expression before the model is evaluated.

Applied per-iteration by spectrafit-solver::lm_problem::set_free_and_tied — the single set_params entry shared by both the nalgebra-LM and faer trust-region front-ends. The FD Jacobian re-applies ties per perturbation, and the analytic Jacobian is swapped for FD when ties are present so chain-rule terms are captured. End-to-end coverage: dispatch::tests::test_tied_amplitude_fit_recovers_ratio and test_tied_fit_reduces_free_param_count in the solver crate.

dataset_offsets: Vec<usize>

Per-dataset point boundaries for simultaneous multi-dataset (“global analysis”) fits: cumulative offsets of length n_datasets + 1, so dataset i owns the concatenated point-range [offsets[i], offsets[i+1]).

Empty by default (single-dataset / fully-global fits). The solver dispatch fills it from the dataset sizes after compile(). The executor only consults it when it is non-empty AND at least one node carries a [NodeEntry::dataset_index]; otherwise every node contributes to all points exactly as before (the all-global path is byte-identical).

Implementations

impl CompiledGraph

pub fn compile(graph: &FitGraphSpec) -> Result<Self, CoreError>

Compile a FitGraphSpec into a CompiledGraph.

Errors

Returns CoreError::Eval if:

• an unknown model type is encountered
• a required parameter is missing from the spec
• two nodes share the same id (would silently corrupt the free-column layout and overwrite components)
• expr_edges contain duplicate targets (cycle / conflict)

pub fn node_params( &self, node_idx: usize, flat: &HashMap<String, f64>, ) -> Result<Vec<f64>, CoreError>

Extract the parameter value vector for node at node_idx from the flat dict.

Values are ordered to match model.param_names().

Errors

Returns CoreError::Eval (via GraphError::MissingParamKey) if flat has no entry for one of the node’s "node_id.param_name" keys.

pub fn n_dims(&self) -> Result<usize, CoreError>

The common coordinate dimensionality shared by every node in the graph.

The executor lays the flat x buffer out point-major (stride = n_dims), so all nodes must agree on how many coordinate components a single point carries. A graph that mixes a 1-D and a 2-D model over the same coordinate grid is rejected.

Returns 1 for an empty graph (degenerate but harmless; evaluation produces zeros).

Errors

Returns CoreError::Eval if nodes declare differing n_dims.