Bug summary

Since 3.11, text is shaped with libraqm, but the paragraph direction is always auto-detected from
the first strong character. An Urdu (or Arabic, Hebrew) sentence that starts with a Latin word, a
product name or a version number is therefore laid out left-to-right, so its words and punctuation
come out in the wrong order. There is currently no way to set the direction.

Code for reproduction

import matplotlib.pyplot as plt

text = "Python اور matplotlib میں اردو؟"  # Urdu sentence that starts with an English word

fig = plt.figure(figsize=(6, 1.6))
fig.text(0.95, 0.70, text, ha="right", fontsize=18, language="ur")
fig.text(0.95, 0.25, "‏" + text, ha="right", fontsize=18, language="ur")  # RLM prefix
fig.savefig("repro_direction.png", dpi=100)

Actual outcome

The top line is laid out left-to-right: "Python" is on the left and the Urdu question mark ؟ ends up
in the middle of the line instead of at the end of the sentence.

Expected outcome

The bottom line, where the only change is a U+200F RIGHT-TO-LEFT MARK prefix: the sentence reads
right-to-left, "Python" is on the right where the sentence starts, and ؟ is at the end.

Additional information

• FT2Font::layout in src/ft2font.cpp sets the text, face, features and language on the raqm
  object but never calls raqm_set_par_direction, so libraqm uses RAQM_DIRECTION_DEFAULT. The
  libraqm documentation for that function notes that this default "has problems when a mainly
  right-to-left paragraph starts with a left-to-right character and vice versa as the detected
  paragraph direction will be the wrong one". The same is true on main.
• Passing language="ur" does not change the direction.
• The reverse also happens: a left-to-right label whose first letter is Urdu is laid out
  right-to-left.
• Workaround: prefix the string with U+200F (RLM) or U+061C (ALM), or wrap it in U+2067 ... U+2069
  (RLI ... PDI).
• This is common in Urdu text: technical terms, names and numbers are often written in Latin script
  at the start of a sentence.
• A second image shows the same sentence in Noto Nastaliq Urdu, as is, with each direction mark,
  and an Urdu-first sentence for comparison.

• Possible fix: a direction text property ("auto" by default, "rtl", "ltr"), passed down to
  raqm_set_par_direction the same way language is passed to raqm_set_language. I would be
  happy to work on a PR if that approach sounds acceptable.

Operating system

Windows 11

Matplotlib Version

3.11.2

Matplotlib Backend

Agg

Python version

3.12.15

Jupyter version

No response

Installation

pip