Block-orientation coarse proposals
Requires 0.1.4 or later; see installation.
Use strategy=:coarse, coarse_groups=[[1,2],[3,4]] for one coordinated proposal from a frozen complete incumbent. Each group defines one macrovariable, so two groups representing eight original variables need child capacity two. Indices are positions in decomposition.labels, after MOI-fixed variables are removed. Groups must be nonempty, disjoint, and in range; their count must fit max_variables. They may cover a subset, including isolates. The complement stays frozen. Group order defines macrovariable order; indices inside groups are sorted. Setters and getters own copies. An empty original model accepts only an empty plan.
For binary variables, the map is xᵢ=rᵢ+(1−2rᵢ)y_g(i) with binary y. For spins, it is xᵢ=rᵢz_g(i) with spin z. The zero binary or all-plus spin macrostate reproduces r; a macro flip reverses every member of its group. Substitute this map into the entire original objective, retaining offset, scale and sense, linear terms, internal interactions, cross-group interactions and interactions with the fixed complement. Normalize a shared-group square using y²=y or z²=1. Consequently E_coarse(m)=E_original(lift(m)) algebraically. Floating-point construction can round; independently recomputed original energy decides every strict improvement. Child-reported objective values are only checked for finiteness. A public OPTIMAL child is also checked against the feasible no-flip macro point: its best returned macro objective cannot be worse beyond atol=1e-12, rtol=1e-12. This check sums exact rationals of the Float64 coefficients actually exported (including coefficient-wise scaling); original-energy arithmetic can round differently. It is internal validation work under the deadline and validation/reconstruction timer, without an additional original-candidate charge. It detects a contradictory child claim and supplies no original global proof.
This opt-in strategy runs even when the whole original model fits capacity. Planning and validation precede constant/zero-scale shortcuts. It accepts serial execution and default selection; component executors, partition/consensus plans, separators, pruning and nondefault selection are rejected. It performs one proposal per optimize!, with no automatic grouping or recursive hierarchy. All maps, starts, groups and counters rebuild on each invocation.
Every result row must be valid and scanned before a strict improvement commits. Malformed output, exceptions, incomplete scans or an expired deadline discard staged candidates. A normally returned INTERRUPTED may supply a fully scanned strict improvement, then stops the parent with INTERRUPTED. Other child limit statuses remain visible. Parent caps and deadlines apply to planning, construction, copying, child work, lifting and original evaluations; exact final parent allowances retain their usual ITERATION_LIMIT precedence. No allowance restarts within this route.
The method is heuristic for the original nonconstant problem. A public OPTIMAL child proves only optimality over the represented assignments. Even singleton groups covering all variables conservatively report LOCALLY_SOLVED in this first implementation: certificate transfer through coefficient arithmetic is not implemented. Original constant objectives retain the existing direct OPTIMAL proof. No global bound or approximation guarantee is supplied.
decomposition.coarse records macro_groups, macro_count, represented_original_count, original_to_macro (zero for uncovered indices), frozen_incumbent, its version and invocation, the affine orientation constants and coefficients, child status and completion/acceptance. Coarse calls record the actual macro source-to-child index map. They do not fabricate an original-to-reduced bijection. Per-call exact describes only the child; original_proof=false states the coarse route's conservative certificate boundary.
The offline example uses nonzero linears and a joint two-group improvement missed by separate exact partition updates. See the bounded comparison for its finite evaluation scope. This feature does not compose preprocessing, sample fixing, consensus or voting into a new production pipeline.