MINGLE: O(E log E) via vendored nanoflann kd-tree

[schochastics]

Follow-up to Phase 6 (#24): make MINGLE scale near-linearly, as the paper intends.

What

• Vendored src/nanoflann.hpp — a header-only kd-tree (nanoflann 1.10.1, BSD). No DESCRIPTION dependency (LinkingTo not needed; it's a self-contained header). Copyright holder added to Authors@R as cph; the BSD header is kept intact.
• mingle_iter rewritten: build the kNN proximity graph over the edges (points in canonical 4-D endpoint space) once with the kd-tree, then coarsen that candidate graph combinatorially each level — instead of the previous O(E²) all-pairs rescan per level. Overall O(E log E).
• New k argument on edge_bundle_mingle() (merge candidates per edge, default 10).

Behaviour preservation

The merge criterion is unchanged (nearest candidate by meeting-point distance → greedy ink-reducing match). With k ≥ edge count the candidate graph is complete, so merges are identical to the exhaustive version — the Phase-6 tests pass unchanged (small-graph mid-spread still 0.5).

Verification

• Full suite FAIL 0 | PASS 82.
• Small graph output identical to the brute-force version.
• 300-edge random graph bundles in ~0.01s; k=3 and k=all both run and produce finite output.

🤖 Generated with Claude Code