Correctness fixes from a brute-force review (1/4)

[Autoplectic]

First of four PRs from a second correctness review. Every fix was reproduced against a brute-force reference implementation or a closed-form value and has a regression test. CHANGELOG.md (Unreleased, 0.4.0) has the full list.

Breaking changes

• correction="bonferroni" is the default for every CSSR learner. Uncorrected CSSR found spurious states in about 10–15% of long-sample runs.
• viterbi returns the n+1 states X_0..X_n, aligned with smooth.
• "crypticity" means C_mu − E everywhere; C± − E is now "bidirectional_crypticity" / bidirectional_crypticity().
• The Gács–Körner model is documented as the common-information variable, not a generator of the process.
• atoms() includes the negative atom, and left_quotients includes the empty residual.
• Brzozowski minimization returns a trimmed DFA, like Hopcroft and Moore.
• Examples emit string symbols.
• There is one golden mean, in the Lind & Marcus form that forbids 11; duplicate examples are removed.
• butterfly_process now matches its paper (the old construction was i.i.d.).

Notable fixes

• Shift synchronization methods crashed with AttributeError.
• Cryptic order was wrong for zero-crypticity processes.
• Bayesian posterior machines started in the data's final state.
• Büchi lasso acceptance missed accepting states visited mid-loop.
• Wheeler order detection, index and minimization were wrong in several cases.
• Modular VPA minimization crashed on valid inputs.
• Language operations on empty languages failed.
• IDFA enumeration undercounted.
• Stack MLE was biased.
• Spectral learning hung on sampled data.
• log_word_probability underflowed on long words.
• Channel statistical complexity depended on the hash seed.
• golden_mean_ghmm did not reproduce the golden mean.
• Topological entropy was inflated on non-right-resolving presentations.
• YAML dropped unused alphabet symbols and could not store sympy values.

Testing infrastructure

• New Hypothesis strategies in sofic.testing.
• Brute-force reference implementations in tests/oracles.py.
• ci and nightly Hypothesis profiles.

Checks

• 1352 tests pass.
• ruff is clean.
• A fresh Sphinx build passes with -W.