Ordered by what unblocks the most, not by size. Every item names the work that owns it, so this page cannot drift away from the backlog.
Pairwise contact construction is gone in all four engines, but constraint-heavy candidates still rebuild the graph. Retaining state plus a candidate delta has to be proven equal to the from-scratch graph.
Measure Python, PHP, Rust and the forced JavaScript path separately on one pinned architecture, reporting cold and warm solve time, throughput, peak memory and uncertainty. Until that lands there is no speed claim.
After a global container-set beam, so optimal becomes a proof rather than a label. 3D packing stays NP-hard; a truthful bounded-search gap is worth more than a promise of exactness.
The contract, harness and registry have landed with a worked connector for a carrier that does not exist. What remains is connectors for ones that do, and the live rate-card ingestion behind them.
shape_type is reserved in the schema and rejected everywhere, so a caller who never opts in is unaffected forever. Convex hulls and compressible goods follow after 1.0.
Rust and the JavaScript fallback are held to validity and an objective floor, not to the reference engines' quality. Closing that gap matters more than any new capability.
The sequence exists so the gate can see the change. Skipping a step does not make the work smaller, it makes it unverifiable.
The suite is pre-1.0: the API is not frozen, coverage and the performance audit are open, and the known gaps are published by name rather than discovered by users.
See where to help ↗