DrudgeCAS / gristmill / 31931519901

92%

master: 92%

Build Type

Pull #44

github

Committed by chenpeizhi

Commit Message

test: Pin that an unrelated term does not disturb the rest

A term sharing no tensor with the others cannot join any biclique with them,
so optimizing them together should cost what optimizing them apart costs,
plus the one addition that puts the two scalars together.  It did not.
Vertices were ordered by a number handed out as edges arrived, so the extra
term shifted the numbers of every other vertex, and the search read them.

On master the three terms alone give `4 n^2 + 4 n`, and in the presence of
the unrelated term they give `2 n^2 + 4 n`: a term that shares nothing, by
luck, buying a better factorization.  The new test fails there by exactly
that `2 n^2`, under OPT, TRAV and EXHAUST alike.

This is the property that content ordering buys over a deterministic
numbering, which is what #39 already gave.  The search only ever compares two
vertices, and under content that comparison is settled by those two alone,
while under numbering it depended on everything inserted before both.  Say
so where the ranks are built, since the code said only that the numbering is
an artefact, not what follows from replacing it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Pull Request Pull Request #44: fix: Stop vertex numbering deciding which constrictions are found

Coverage Stats

90 of 90 new or added lines in 1 file covered. (100.0%)

37 existing lines in 1 file now uncovered.

1854 of 2025 relevant lines covered (91.56%)

0.92 hits per line