Bounded-width VE canonicalizer (representation preprocessing)

[isPANN]

Summary

Adds a static, width-aware constraint-network canonicalizer — bounded-width variable elimination — as a preprocessing step, plus the supporting selector and representation widening. Eliminating structurally-cheap variables before search shrinks the branch set and gives a several-fold end-to-end speedup on factoring (4.5–9.2× median at 16–20 bit), with answers read straight off the reduced network.

What's in src/

• preprocessing/canonicalize.jl — bounded_ve_canonicalize(cn; budget_B, protected, order=:weighted_min_fill). Generalizes precontract_degree2! from "eliminate any degree-2 variable" to "eliminate any variable whose elimination width stays ≤ budget_B, in weighted-min-fill order". budget_B is the single fine↔coarse knob. Unlike degree-2-greedy (which raised treewidth), min-fill ordering with a per-step width cap cuts the treewidth proxy 50–70%.
• Protected read-out variables: protected variables (e.g. factor bits) are never eliminated, so they survive into the reduced network and their values are read directly via result.orig_to_new — no variable-elimination back-substitution.
• DiffLookaheadSelector (branch_table/selector.jl): difficulty-guided lookahead selector that cuts the search tree vs MostOccurrenceSelector.
• TensorData config indices widened UInt16 → UInt32 (core/static.jl, branching/propagate.jl, branch_table/contraction.jl): lifts the per-tensor arity cap from 16 to 32 so the canonicalizer can use larger budgets.

budget_B is a plain hyperparameter.

Tests

• New test/canonicalize.jl: canonicalizing a factoring instance preserves satisfiability, keeps the protected factor-bit variables alive, shrinks the branch set, and the factors are read off the solution (a*b == N). 16/16 pass. Runs early in runtests.jl so the suite reaches it before the cadical-dependent solve_factoring test.
• The only local suite error is the pre-existing solve_factoring cadical path (missing libcadical_mine.dylib), unrelated to this branch.

🤖 Generated with Claude Code