Step 2 — Build the partial graph

A bookkeeping step. Given the lagged structure from Step 1:

1. Add an undirected contemporaneous edge between every pair (i, j) of variables. This is the fully-connected starting skeleton; Step 3 will thin it.
2. Tentatively mark every variable as a changing module. Step 3 will prune via CI tests against the surrogate.

No statistical tests are run here. Lagged edges are untouched.

Why this is its own step

Two reasons:

1. Separation of concerns. Steps 1 and 3 are statistical tests; Step 2 is pure data-structure manipulation. Keeping it separate makes the boundary between "what we observed" and "what we initialize" explicit.
2. Easier to swap in different starting structures. A user with prior knowledge can replace build_partial_graph with their own initializer (e.g., one that excludes physically impossible contemporaneous edges) without touching Step 3 or Step 4.

API

build_partial_graph

build_partial_graph(graph: TimeSeriesGraph) -> TimeSeriesGraph

Add the fully-connected contemporaneous skeleton to graph.

Modifies graph in place and returns it. The lagged edges already present are preserved untouched.

Parameters:

Name	Type	Description	Default
graph	TimeSeriesGraph	The output of Step 1 (lagged adjacency identification).	required

Returns:

Type	Description
TimeSeriesGraph	The same graph with all contemporaneous pairs (i, j) for i < j connected by an undirected edge.