[Autoplectic]

Re-lands #35, which was merged into dit-channel-adapter after #34 had already merged, so it never reached main. Same five commits, cherry-picked onto current main, including a ty fix for #34's test imports. Needs dit's feedback codes (dit/dit#304).

Summary

• FiniteStateChannel feedback methods: feedback_initial / feedback_step step the channel one use at a time, with the normalized forward vector as the transmitter's belief. feedback_capacity brackets the capacity with feedback. dit's feedback codes now run over Mealy-machine channels.
• sofic.generators.trapdoor_code.trapdoor_code(k): the zero-error feedback code of Permuter, Cuff, Van Roy & Weissman (2008, Sec. VII) for the trapdoor channel.
  • Each input is chosen relative to the ball in the channel (same or opposite), which the encoder knows from the feedback.
  • Messages are enumerated into words with no two "opposite" choices in a row; the decoder reads them backward from whether consecutive outputs differ.
  • Each transmission starts by flushing the unknown initial state, 3.5 uses on average.
  • The module is not imported by sofic.generators, so sofic still imports with older dit.
• Results on trapdoor():
  • Horstein's scheme reaches about 0.60 / 0.62 / 0.63 bits per use for k = 8 / 10 / 12, already above the best no-feedback upper bound of 0.584 (Huleihel et al. 2021).
  • The trapdoor code is exactly zero-error with E[N] = N + 3.5 and reaches about 0.663 at k = 64, approaching log2(phi) ≈ 0.694.
• Docs: a "Codes with feedback" section in docs/generators/channel_coding.rst, a Horstein1963 bib entry, and changelog entries.

Test plan

• tests/test_channel_feedback.py: 12 tests. They check that feedback_step reproduces log_likelihood (hypothesis), that a one-state channel matches dit's BSC, that exact matches Monte Carlo on the trapdoor, and that the feedback-capacity bracket contains log2(phi). Horstein must beat 0.584. The enumerative code must be a bijection (hypothesis), and the trapdoor code must be exactly zero-error with E[N] = N + 3.5. The tests skip when the installed dit lacks feedback codes.
• ruff check / ruff format are clean.