TurncatedNormal gives wrong results sometimes

Truncated normal gives wrong values sometimes. Seems to be when the scale is relatively small, but in surprising situations where you'd expect it to work like TruncatedNormal(1, 0.1, 0, 10).

MVCE

import jax
import jax.numpy as jnp
import pytest
import tensorflow_probability.substrates.jax as tfp

tfpd = tfp.distributions


@pytest.mark.parametrize("scale", [0.01, 0.1])
@pytest.mark.parametrize("low", [0.0, 0.])
@pytest.mark.parametrize("high", [10, jnp.inf])
def test_truncated_normal(low, high, scale):
    dist = tfpd.TruncatedNormal(1.0, scale, low=low, high=high)
    u = jnp.linspace(0., 1., 100)

    samples = jax.vmap(dist.quantile)(u)
    assert jnp.all(samples >= low)
    assert jnp.all(samples <= high)

Reproduce the MVCE with the parametrized scale, low, and high values, then trace the tfpd.TruncatedNormal quantile path for the small-scale cases. Compare the returned samples with the bounds and establish a regression test that captures the correct behavior for these inputs.