feat(math): add graph morphism operations (#1856)

[morluto]

Summary

Implements graph morphism operations for homomorphism checking, search, core detection, and retraction.

Closes #1856

Operations

Operation	Description
graph.homomorphism.check	Check if a vertex map is a graph homomorphism
graph.homomorphism.find	Find a homomorphism via backtracking search
graph.core.check	Check if a graph is a core (no non-injective endomorphism)
graph.retraction.check	Check if a retraction onto an induced subgraph exists

Design

• Exact adjacency checking: All homomorphism checks use exact edge-preserving verification over adjacency sets.
• Backtracking search: Homomorphism finding uses systematic backtracking over all possible vertex maps, bounded by the vertex count limit (≤64).
• Core detection: A graph is a core iff it has no non-injective endomorphism; the operation searches for one.
• Retraction: Checks for a homomorphism from G to an induced subgraph H that fixes every vertex of H.

Continue this on Linzumi

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[morluto]

Deep review summary

Verdict: REQUEST CHANGES — the core checker prunes every partial assignment incorrectly and reports many non-core graphs as cores; all search operations are also effectively unbounded.

1. graph.core.check is wrong for any ordinary graph with edges

During backtracking, _is_endomorphism() checks all source edges even though most vertices are still mapped to -1. At the first partial assignment, an edge touching an unassigned vertex becomes (mapped, -1) or (-1, -1), is absent from the adjacency set, and the branch is rejected before the remaining vertices can be assigned.

Thus the search generally never reaches a complete mapping and reports is_core=True. A path on three vertices is a concrete counterexample: it has the non-injective endomorphism mapping both endpoints to one endpoint and the middle vertex to the other, so it is not a core, but this implementation reports it as a core.

Use the same partial-edge discipline as the homomorphism finder: test an edge only after both endpoints are assigned, and replay the full homomorphism plus non-injectivity predicate at the leaf. Add regressions for P3 (not a core), K_n (core), bipartite graphs retracting to K2, and edgeless graphs.

2. The accepted search domain is not bounded

Homomorphism search can explore |V(H)|^|V(G)|, core search n^n, and retraction search |S|^(n-|S|), with public vertex bounds of 64 and no node budget, timeout status, or complexity-derived request restriction. Those are not total bounded operations.

Either impose small, justified work-product bounds and test worst cases, or expose max_nodes with explicit FOUND / NOT_FOUND_COMPLETE / INCOMPLETE outcomes. A timed-out search must not be reported as mathematical nonexistence.

The direct graph.homomorphism.check predicate is correct for the represented loopless simple graphs.