HGraphML v0.1.0: Trainable Quotient-Tower Graph Message Passing

HGraphML v0.1.0 is the first lightweight public research release.

This release establishes the first executable bridge from state_collapser quotient towers into trainable graph message passing. A known graph is treated as already discovered, collapsed through a state_collapser partition tower, used for coarse message passing, and then lifted back over node and edge fibers to produce fine-graph readouts inside a PyTorch computation.

Highlights:

• Adds the TensorGraph surface for known directed graphs.
• Adds a direct adapter from graph data into state_collapser quotient towers.
• Adds node-fiber and edge-fiber readouts from tower tiers.
• Adds uniform pullback, fiber-normalized, and learned lift operators.
• Adds message containers, edge-message MLPs, pooling, and readout helpers.
• Adds collapse_messages(...) as the package-native orchestration call.
• Adds a tiny supervised train-step helper and learned-lift demo.
• Adds tests, typing, Ruff linting, CI, build metadata, usage docs, API notes, design docs, and contributor guidance.
• Pins the upstream state_collapser dependency to the public v0.6.0 research baseline.

This release does not claim graph-ML speed-up. It demonstrates that quotient-tower-backed graph message passing is executable, differentiable, and package-shaped. Serious benchmarking is the next major milestone before any performance claims.