Add spectral reconstruction from sequences to ε-machines

[Autoplectic]

Summary

• Wire Hankel-SVD spectral learning through mixed-state extraction so EpsilonMachine.from_sequence(..., method="spectral") returns an ε-machine.
• Use a Mealy projection when the learned operators are non-negative; otherwise enumerate mixed states of the signed operators (Ellison et al. 2009) instead of clustering.
• Document the path and drop spectral from the “not yet implemented” reconstruction list. transCSSR is already in sofic as ε-transducers; conjugate BSI, subtree merging, and mixed-state extraction cover the algorithms sometimes labelled CSM / Gibbs BSI / NSD.

Test plan

• uv run python -m pytest tests/test_epsilon_inference.py tests/test_spectral_inference.py -q --numprocesses=0
• Golden mean and even process recover two states from exact block probabilities
• Bernoulli from_sequence(..., method="spectral") recovers a single state
• Signed-operator extraction raises SpectralInferenceError when max_states is exceeded

Made with Cursor