CSSR and transCSSR learn from independent Trials

[Autoplectic]

Summary

• Adds sofic.inference.Trials. When learn_epsilon_machine_cssr, learn_epsilon_transducer_cssr, and suggest_max_history get a Trials, they pool counts across trials. No history spans the boundary between two trials. A dit.inference.Trials is accepted too.
• Adds SuffixCounts.from_sequences, and JointSuffixCounts.from_sequences now accepts paired Trials. A single sequence behaves exactly as before.
• Documents transCSSR's Mealy convention: the transducer reads x_t, then emits y_t. To model a channel with delay d, shift the input by d.

Needed by infoflow's machine layer (dit/infoflow machines-layer).

Tests

• New tests/test_cssr_trials.py:
  • Hypothesis property tests: pooled counts equal the per-trial sums, and no joint history spans two trials.
  • Learning from split trials gives the same topology as the unsplit sample (golden mean, even process, a binary channel).
• The existing CSSR, transCSSR, and inference suites pass.

Made with Cursor