GeezerVison

Summary

CPython's float.__round__ (Objects/floatobject.c) routes through _Py_dg_dtoa and rounds at the decimal level. The previous round_float_digits multiplied by 10**ndigits and rounded at the IEEE 754 binary level, which diverges for values that aren't exactly representable.

>>> round(2.675, 2)
# CPython 3.14.4: 2.67
# RustPython main: 2.68 ❌

2.675 stores as 2.67499...; dtoa correctly rounds it down to 2.67, but (2.675 * 100.0).round() / 100.0 lands on 2.68 because the multiplication produces a phantom 267.5 tie that round-half-to-even snaps up.

Rust's {:.*} float formatting uses dtoa-style algorithms (Grisu3 + Dragon4 fallback) and matches CPython's _Py_dg_dtoa byte-for-byte. Replacing the multiply-then-round step with format!("{:.*}", ndigits, x).parse() removes the entire phantom-tie class for ndigits >= 0.

For ndigits < 0 the existing divide-then-round path is preserved because dividing typical inputs by 10**|ndigits| produces genuine half-integer ties rather than synthesizing them.

Verification (CPython 3.14.4)

Tests unmasked

Lib/test/test_float.py::RoundTestCase:

• test_matches_float_format — rounds against float(format(x, '.{n}f')) across many values; this is exactly what the fix does internally.
• test_previous_round_bugs — explicit "round-half-even" cases (e.g. round(25.0, -1) == 20.0).

Test runs

cargo run --release -- -m test test_float test_builtin test_decimal \
    test_fractions test_complex test_math test_format test_long test_int \
    test_statistics test_re

1573 tests pass, 0 regressions. All 188 extra_tests/snippets/*.py pass under the CI feature set (stdlib,importlib,stdio,encodings,sqlite,ssl-rustls,host_env).

A regression block is added to extra_tests/snippets/builtin_round.py.

Summary by CodeRabbit

• Bug Fixes

  • Improved rounding to more closely match CPython’s behavior: better handling of halfway ties (banker’s rounding), negative ndigits, signed zeros, and non-finite inputs.

• Tests

  • Expanded rounding tests to cover banker's rounding, negative ndigits, exact-half cases, NaN/infinity behavior, signed zero, and out-of-range ndigits.