[MilesCranmerBot]

Adds an optional Rust backend, PySRRegressor(backend="rust"), which runs the search with symbolic_regression.rs through its Python package symbolic-regression-rs. The Julia backend stays the default, and nothing changes for existing users.

This depends on the Python package landing in symbolic_regression.rs (astroautomata/symbolic_regression.rs#57) and being published to PyPI as symbolic-regression-rs 0.1.0. Until then, CI builds it from a pinned commit.

What changes

• backend="julia" (default) or "rust". A new pysr[rust] extra installs symbolic-regression-rs.
• Importing is unchanged by default: import pysr still loads Julia exactly as before, and backend="rust" works with that normal import. Setting PYSR_BACKEND=rust before import pysr skips Julia entirely (juliacall is never imported, so Julia is never installed or started), makes "rust" the default backend, and makes Julia-only objects such as pysr.jl raise an error that says to unset the variable.
• With the Rust backend, the regressor's settings are translated to SymbolicRegression.jl's option names and passed to the Rust engine. Parameters that only make sense in Julia (loss_function, loss_function_expression, complexity_mapping, turbo, bumper, autodiff_backend, logger_spec, cluster_manager/procs and related multiprocessing options, fast_cycle, type specs, precision=16, custom Julia mutations or plugins) raise ValueError naming the parameter when set to non-default values. The Rust engine takes real float32/float64 data.
• Results go through the existing code: the engine writes SymbolicRegression.jl-format CSV files, so equations_ has the same columns and dtypes, scores, SymPy/JAX/PyTorch exports and from_file behavior as the Julia path. Ordinary expressions predict through the same SymPy-lambdified functions; template expressions predict with the engine's evaluator, since their equation strings are not SymPy.
• warm_start resumes from the engine's state, kept in rust_state_stream_; pickled regressors predict after unpickling.
• Separate fix: from_file now loads CSV-only multi-output runs, which raised FileNotFoundError before (first commit, with a test).
• Docs: the backend choice, installation, PYSR_BACKEND, and the options the Rust backend does not support.
• CI: a new job installs Rust, builds the engine's Python package, and runs the Rust-backend tests with and without PYSR_BACKEND=rust.

Verification

• Rust-backend tests: 61 passed with the normal import; 59 passed and 2 skipped (Julia-only checks) with PYSR_BACKEND=rust. They cover fits and predictions, equations_ columns and dtypes against the Julia path, exports, multi-output, units, guesses, templates, string losses and early stops, warm start, pickling, options taking effect (operator complexity, nested constraints, seeds, maxsize), errors for Julia-only parameters, and the missing-package message.
• The existing test suite with the Julia backend: 189 tests, OK.
• In a fresh environment with only PySR and the Rust package installed, PYSR_BACKEND=rust imports every PySR module, fits and predicts without importing juliacall, and never creates a Julia installation.
• With PYSR_BACKEND unset, an independent check found the import order, environment variables, __all__, and pysr.jl unchanged.

[codecov]

Codecov Report

❌ Patch coverage is 38.42365% with 125 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pysr/rust_backend.py	0.00%	89 Missing ⚠️
pysr/julia_import.py	59.18%	20 Missing ⚠️
pysr/sr.py	65.62%	11 Missing ⚠️
pysr/__init__.py	80.00%	2 Missing ⚠️
pysr/julia_helpers.py	81.81%	2 Missing ⚠️
pysr/_backend.py	85.71%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!