fix: Fix flaky matplotlib-to-plotly conversion of touching bars (negative bargap)

[robertoffmoura]

mpl_to_plotly intermittently crashes when converting figures with touching bars (e.g. plt.hist). The bar gap is computed as bar_start[i+1] - bar_end[i], and for adjacent bars floating-point noise can make this a tiny negative number (e.g. -8.88e-16). Plotly's bargap property only accepts values in [0, 1], so the conversion fails with:

ValueError: Invalid value of type 'numpy.float64' received for the 'bargap' property of layout
    Received value: np.float64(-8.881784197001252e-16)

Reproduces with:

import warnings
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import plotly.tools as tls

np.random.seed(0)
plt.hist(np.random.randn(10000), 30)

with warnings.catch_warnings():
    warnings.simplefilter("ignore")
    tls.mpl_to_plotly(plt.gcf())

Fix: get_bar_gap now clamps the gap to [0, 1], so float noise around zero maps to bargap = 0 (touching bars) and genuine gaps are unchanged.

[camdecoster]

Could you please make a couple of changes and add a CHANGELOG entry?

[camdecoster]

We need to normalize gap0 to calculate the correct bargap.

Suggested change

	# plotly's bargap must be in [0, 1]; clamp to guard against
	# floating point noise (e.g. -8.9e-16 for touching bars)
	return min(max(gap0, 0.0), 1.0)
	# Plotly's `bargap` is a fraction of the distance between bar positions, not a gap in data units
	# so we need to normalize `gap0` with `bar_delta`
	bar_delta = bar_starts[1] - bar_starts[0]
	if bar_delta <= 0:
	return None
	# Clamp to guard against floating point noise, such as -8.9e-16 for touching bars
	return min(max(gap0 / bar_delta, 0.0), 1.0)

[camdecoster]

Let's make this test deterministic:

Suggested change

	fig, ax = plt.subplots()
	ax.hist(np.random.randn(1000), 30)
	plotly_fig = tls.mpl_to_plotly(fig)
	assert len(plotly_fig.data) == 1
	assert 0 <= plotly_fig.layout.bargap <= 1
	# Seed 0 makes the first gap slightly negative (-4.4e-16)
	rng = np.random.RandomState(0)
	fig, ax = plt.subplots()
	ax.hist(rng.randn(10000), 30)
	plotly_fig = tls.mpl_to_plotly(fig)
	assert len(plotly_fig.data) == 1
	assert plotly_fig.layout.bargap == 0