Fleet & Topology Optimization¶

CloudSlash does not claim a MILP solver it does not ship. The optimisation surface is three concrete core algorithms:

1. Multi-dimensional vector bin-packing (pkg/engine/tetris)¶

The Tetris packer fits workloads into nodes across several resource dimensions (CPU, memory, and additional vectors) with cosine-angle matching. Kubernetes CPU/memory packing is the two-dimensional special case. It reports physical bins, utilisation, and the right-sizing recommendation.

2. Interruption and risk prediction (pkg/engine/oracle)¶

• HMM with Viterbi / forward / Baum-Welch for spot-interruption regimes.
• Bayesian zone risk with decay.
• Conformal prediction intervals and a Thompson-sampling pool allocator.

These are forecasting models, not exact optimisers. Outputs carry uncertainty.

3. Min-cut topology (pkg/graph/flow)¶

Dinic's algorithm computes maximum flow and the minimum cut over a capacitated graph. Use it to find the cheapest set of edges that separates a source side from a sink side — egress anchors and blast-radius bottlenecks. The solver returns the cut value, the reachable partition, and the crossing edges.

n := flow.New(vertices)
n.AddEdge(from, to, capacityUSDPerGB)
value, crossings := n.MinCutEdges(source, sink)

All three run in core. The CLI, TUI, and Horizon surfaces only display their results; none reimplement the algorithms.