The wall framing takeoff orders studs and plates with a column-generation cutting-stock solver. Today's runs made it several times faster (warm starts, one knapsack table, batched rows, an approximate dual steepest-edge rule). Each change left the answer within a fraction of a percent of the LP lower bound, but the pool of cutting patterns differs, so the rounding lands on a different integer order.
On the small built-in sample house that showed: the lumber line moved from $2,389.60 to $2,384.44 (238 to 236 2x4 boards). The committed sample-bid-package.html still held the older figure, so sample-pack.test.mjs failed on main. I first read the assertion backwards and suspected a regression; the order of assert.equal(actual, expected) arguments said the stale file was the "actual". Tracing the total through each engine commit showed the cheaper order is the current one.
Fix: regenerate the sample (node build-sample-pack.mjs). Lesson: a pinned generated file belongs to the solver's output, and a solver speedup that preserves quality can still change it. Invented house and prices only.