M1 solver reconnaissance

Date: 2026-07-30 Solver: microlp = "=0.6.0" Command: cargo run --release --locked --bin reconnaissance

The measurement ran while another long-running optimization campaign shared the machine. The values establish horizon order of magnitude and API suitability; they are not isolated cross-solver benchmarks.

API findings

• Bounded continuous variables, equality/inequality rows, minimization, objective values, primal values, termination reasons, and statistics are available.
• Stats::lp_iterations reports cumulative simplex pivots.
• Problem and SolveOutcome satisfy Send; one rebuilt problem can be solved independently per candidate worker.
• Solution supports adding a row and fixing/unfixing a variable, but the public API cannot replace arbitrary capacity bounds or constraint right-hand sides. Candidate dispatch models are rebuilt rather than falsely advertised as warm-started.
• Pure LP solves have a wall-clock limit but no public simplex-iteration limit. The MIP node limit is irrelevant here. A wall-clock limit would make outer fitness host-dependent, so the tutorial uses unlimited solves and accepts only proven optima.

Gate decision

microlp remains the only dispatch solver and preserves a pure-Rust hot path. The main optimizer campaigns use 96-step smoke and 288-step publication models. Seasonal H₂ sizing is a focused 1,460-period, six-hour chronological arm; its selected design receives an independent 8,760-hour replay.