The pick path optimizer finds the shortest walk a picker can take through a warehouse with parallel aisles. It uses the exact dynamic programme of Ratliff and Rosenthal (1983): walk the aisles left to right and remember, for the front and back corner of the current aisle, only whether each has odd or even degree and whether the two belong to the same piece of the tour so far. That gives a handful of states per aisle, so the solver is linear in the number of aisles.
To batch orders into carts, the tool scores thousands of candidate carts with that exact router. A profile of its tests showed most of the time going on the bookkeeping inside the DP, not the search. For every state, in-aisle pattern and pair of cross-aisle copies, it built four small objects and a string key, then relabelled the pieces so equivalent states compared equal.
The result of each step depends only on the state and a few bits of the pattern. Across a whole benchmark of 16,000 random pick lists there are 396 distinct combinations. So the solver now works them out once per process, keeps states as small integers, and looks the answer up. It also skips building the actual walk when the caller only wants its length, which is what the cart search asks for.
The answers are unchanged. 1.8 MB of routes, rankings and cart assignments from a fixed-seed script match byte for byte, as does the benchmark table, and a new test pins the length-only path to the full one on 1,200 random layouts. The tool's test suite went from 6.3 s to 2.1 s. In the browser the same change makes the "compare policies" step faster on large order lists.
Details and what is left: docs/history/ci-perf-2026-11-18-pick-path-step-memo.md.