mip presolve: worklist-fixpoint propagation, singleton-column elimination

[andig]

Summary

Two presolve-side reductions, follow-up to #13:

• Bound propagation runs to fixpoint via a row worklist (mip/presolve.go). The previous 4-full-pass cap both overworked (rescanning untouched rows) and underworked (stopping short of the fixpoint on long implication chains). The worklist seeds all rows and re-enqueues only rows touching a tightened column; the fixpoint is unique since bounds only shrink, with a generous pivot-count safety cap for degenerate chains. Node-level propagation gets cheaper and tighter: golden case 020 6.9s → 6.5s, 016 0.5s → 0.2s, 018 1.2s → 0.9s.

• Singleton-column elimination with exact postsolve (mip/eliminate.go). Costed continuous columns appearing in exactly one row are substituted out before the LP is built: a column whose cost is capped only by its row pins that row at any optimum (x = (b - rest)/a, cost folded onto the row's remaining columns); one whose row never resists its cost is fixed at its bound. Folding can re-classify same-row siblings, so the pass iterates. Solutions, duals, reduced costs and row activities are reconstructed exactly in the caller's column space, including across the warm-restart path.

An honest note on the second one: on the evcc golden instances it finds nothing — their ~616 continuous singletons are penalty slacks whose cost fights the row (a max(0, ·) term no exact linear presolve can remove). An earlier census misread them as eliminable by ignoring ObjSense. The transform does fire on minimize-form models with revenue-style singletons, e.g. anything fed through the cbc CLI, and comes with unit tests covering tight-pin, bound-fix, and partial same-row chains plus a dual-identity check.

Testing

• go test ./... green; three new unit tests for the elimination (structure, values, duals).
• Full 13-case golden gate green; case 020 stable at ~6.4s, trajectories bit-identical where elimination finds nothing.

🤖 Generated with Claude Code