The graph-structured stack itself: nodes keyed by {state, stream_position} (deduplicated -- two derivations that reach the same state at the same position share one node, the actual efficiency win over independent parallel stacks), edges labeled with whatever that edge's own transition produced (a shifted %Grammar.VM.Token{}, or a reduced nonterminal's own already-built captures map) plus the provenance accumulated along it so far.

A node can have more than one incoming edge exactly when two previously-distinct derivations converge (a shift/reduce or reduce/reduce conflict both survived, or two different reduce paths happen to land on the same state and position) -- that's genuine local ambiguity, not a bug, and paths/3's whole job is enumerating every distinct way to walk backward through however many edges a reduce needs to pop, independently for each of a node's incoming edges.

Positions live on the node identity itself ({state, position}), not on the edge -- which is what makes a :text-kind capture's span (Grammar.LRTable.Captures) derivable per RHS position directly from the two nodes an edge connects, with no separate bookkeeping needed.