pytest.approx Decimal repr omits or misstates the effective tolerance

For Decimal values, pytest.approx can display the raw relative factor instead of the effective tolerance, ignore a larger absolute tolerance, or show ??? with valid defaults.

from decimal import Decimal as D

import pytest

for options in (
    {"rel": D("0.01")},
    {"rel": D("0.01"), "abs": D("2")},
    {},
):
    obj = pytest.approx(D("100"), **options)
    print(repr(obj), obj.tolerance, sep=" | ")

Output:

100 ± 1.0e-2 | 1.00
100 ± 1.0e-2 | 2
100 ± ??? | 0.000100

The comparison windows are [99, 101], [98, 102], and [99.9999, 100.0001]: both endpoints compare equal, while 101.1, 102.1, and 100.00011, respectively, compare unequal. The comparisons are correct in these examples; the display does not communicate their tolerance.

Expected: the values after ± should represent offsets of 1, 2, and 0.0001. The float equivalent already displays 100.0 ± 1 for pytest.approx(100.0, rel=0.01).

The existing test_decimal_approx_float_rel explicitly expects the raw relative factor. The cases above additionally demonstrate the omitted dominating absolute tolerance and unknown default. Related #13530/#13543 address a FloatOperation exception during repr.

Reproduced on pytest 9.1.1 and development commit 3fd8675d6d798507c06cf9c60753be6d9d7b0e17, with CPython 3.12.3 on Linux 6.8.0-100-generic x86-64.

Installed packages (uv pip list --offline; this environment has no pip module):

Package          Version
---------------- -------
attrs            26.1.0
hypothesis       6.168.0
iniconfig        2.3.0
numpy            2.5.3
packaging        26.3
pluggy           1.6.0
pygments         2.21.0
pytest           9.1.1
sortedcontainers 2.4.0

GPT-6 Astra assisted; I manually reviewed this report.

Start with testing/python/approx.py and the existing test_decimal_approx_float_rel test, then run the Decimal examples from the report to inspect repr and tolerance values. Add focused coverage for the three Decimal cases and verify that the displayed offsets match the effective comparison windows while existing tests continue to pass.