v0.5.0 - Equivalence budget, minimize default fix, expanded docs

Highlights

• prove_equal(..., max_expressions=N): total-work budget across both bidirectional-BFS frontiers. Un-provable queries that previously exhausted the full depth-bounded reachable set now return None in bounded time. Measured impact: an un-provable boolean case went from 20,174 ms to 363 ms with max_expressions=2000.
• minimize() default flipped to include_unidirectional=True: the old default silently ignored every => rule, producing "no improvement found" on expressions with obvious unidirectional simplifications. Pass include_unidirectional=False to restrict to strict reversible equivalences.
• Experiments directory with runnable benchmarks that validate correctness against theory (n! × Catalan(n-1) class sizes under associativity + commutativity) and measure timing across proof, enumeration, minimize, and random-walk features.
• Expanded docs: new Equivalence & Proof guide, new API reference sections, new worked examples, and a CHANGELOG.md.

Breaking changes

• minimize() default for include_unidirectional changed from False to True. See above.
• OptimizationResult.improvement_ratio semantics changed in 0.4.0; use cost_ratio for the old behavior. (This landed in a prior commit but is worth re-flagging for anyone who skipped 0.4.)

Added

• prove_equal(..., max_expressions=N) work budget.
• experiments/features_benchmark.py (7 probes) and experiments/scaling.py.
• docs/equivalence.md, new API and examples sections, README section for equivalence and optimization.
• CHANGELOG.md.

Fixed

• Removed dead check_intersection() helper in prove_equal().
• Collapsed a dead if/else branch in equivalents().
• Replaced 7 stringly-typed "failed" sentinel checks with wrap_bindings().
• __version__ in rerum/__init__.py is no longer stale.

Tests

449 tests passing (up from 443 in 0.4.0). New test classes: TestProveEqualMaxExpressions, plus new regression tests for the minimize() default.

Install

pip install -U rerum

Full changelog: CHANGELOG.md

🤖 Generated with Claude Code