[Autoplectic]

Summary

_polytope_vertices tried every set of n - 1 rows of the support-type polytope (n + 2(2^n - 1) rows) in Fraction arithmetic. For a 4-state survey support with 6 maximal rectangles that is C(132, 5) ≈ 3.4e8 exact solves: hours, and it is why most 4-state survey runs hit their hard timeout before any time budget could act.

Now:

1. _irredundant drops rows the remaining rows imply. A row is dropped only when an exact LP dual bound (certificates.certified_lp_minimum) proves its maximum is at most its right-hand side; nonnegativity rows are always kept. 261 rows -> 40 for 7 rectangles.
2. Candidate active sets are screened in batched numpy (a scale-free determinant test for nonsingularity, then feasibility to 1e-7). Every true vertex is exactly feasible on a nonsingular set, so it passes; each candidate is then solved and checked exactly as before.

Timings: 6 rectangles, hours -> 0.3 s; 7 rectangles, 7 s. (2, inf): G = 2 - 3/8 log2 3 and the certified flow bound 1.49702 are unchanged.

Test plan

• New tests/test_generator_bounds.py: hypothesis test that the pruned enumeration equals brute force on random sparse 3x3 supports; a 7-rectangle staircase support runs in < 60 s.
• Pruned and brute-force vertex sets are identical on all 16 survey supports with 2–4 rectangles.
• Coupled-generator, phase-bound and pipeline suites; ruff; ty.

Made with Cursor