Reduce connectivity separation and preprocessing overhead

[berendmarkhorst]

Summary

Repeated per-terminal maximum flows, sparse-matrix allocations, quadratic terminal bottleneck storage, and repeated dual ascent were dominating Python-side solver time. This PR reduces that work while keeping maximum-flow fallback and exact-solve certification.

• Screen satisfied connectivity demands with one traversal; use capacity-checked, deduplicated connectivity cuts for integer candidates. Keep minimum cuts in LP separation, including main's nested cuts for remaining demands.
• Skip residual construction for satisfied flows, extract cuts with NumPy, and default separation to one configurable worker. Correct the NetworkX fallback's source partition by completing the flow first.
• Replace the quadratic terminal bottleneck table with a lazy O(T log T) tree-path index and reuse the final unchanged preprocessing ascent once after validating its inputs.
• Use compiled bounded Dijkstra for long-edge reduction where graph size/density justify it and the existing node-search cap cannot bind. Preserve deterministic deletion order.

Integrated with main at 7b4142e, preserving enumeration-safe preprocessing, terminal contractions, the few-terminal DP, cut aging, primal portfolio options, and solve-status handling. No new runtime dependency.

Validation and performance

• Full integrated suite: 1,140 passed, 26 skipped, 20 xpassed, 92% coverage. Independent graph oracles cover cut validity/completeness, indexed bottlenecks, and native long-edge deletions; mutation tests cover ascent reuse.
• Sphinx HTML documentation builds with explicit version metadata overrides for the local environment.
• Current-main measurements and reproduction details: PR_VALIDATION.md. Earlier profiling rounds and raw observations are also committed; their baselines predate recent main improvements.

Latest comparison: 108 solves, all at the known optimum with zero gap. Median total runtimes against main 7b4142e:

Configuration	Main	PR	Speedup
c13, default	4.279 s	2.500 s	1.71×
d08, accelerated	9.079 s	2.700 s	3.36×
d13, accelerated	11.064 s	2.936 s	3.77×
d15, default	1.254 s	0.405 s	3.09×

Changing cut selection can change Gurobi's search tree. The report includes unchanged cases and regressions; local timings do not establish a general speed guarantee or parity with SCIPJack. Residual-flow reuse and further certificate-allocation reductions are left for separately benchmarked work.

AI assistance: Codex assisted with implementation, profiling, tests, documentation, and PR preparation.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 97.90576% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
steinerpy/graph_reducer.py	97.82%	2 Missing ⚠️
steinerpy/mathematical_model.py	96.72%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!