Keep create_text_sprite text through texture atlas rebuilds

[pvcraven]

Problem

create_text_sprite sprites go blank when the texture atlas rebuilds itself.

create_text_sprite makes an empty texture (Texture.create_empty) and draws the text straight into the atlas with render_into. The texture's own image stays blank. When the atlas runs out of room and some images have been freed, DefaultTextureAtlas._add calls rebuild() instead of growing. rebuild() clears the atlas and redraws every texture from its image, so every text sprite becomes a blank image. Growing the atlas with resize() copies pixels on the GPU, so it isn't affected, which is why this only shows up sometimes.

Reproduced directly: a 20pt red "Hello World" sprite has 155,453 total alpha in the atlas before atlas.rebuild(), and 0 after.

This is why #2917's CI fails. test_same_text_different_style (added in #2915) fails on Python 3.10, 3.12 and 3.13 there: the red sprite reads back blank after the blue one is created. In that run, an earlier test left the shared atlas full enough that adding the blue sprite triggered a rebuild. #2917 changes which examples run and how often, which changes how full the atlas is by then. I reproduced exactly that sequence with a small atlas: a full atlas with some images freed, then the red sprite, then the blue one, which rebuilds the atlas. The red sprite is blank without this fix and unchanged with it.

Fix

After drawing, create_text_sprite calls texture_atlas.update_texture_image_from_atlas(texture), the atlas method meant for "render into the atlas and need to update the texture with the new pixel data". The texture's image now holds the text, so a rebuild redraws it correctly. As a bonus, sprite.texture.image now contains the text, for code that reads it.

Cost: the GPU read adds about 0.2–0.6 ms per create_text_sprite call, measured 2 × alternating:

	Before	After
"Score: 100", 16pt	~1.3 ms	~1.5–1.7 ms
a long 40pt title	~1.6 ms	~2.2 ms

create_text_sprite is called once per sprite, not every frame, so this seemed worth it for sprites that don't randomly vanish.

Tests

• New test_survives_atlas_rebuild: calls atlas.rebuild() directly and checks the sprite's pixels are unchanged, and that texture.image matches the atlas. It fails without the fix, whatever state the shared atlas is in.
• The existing flaky test: test_same_text_different_style now passes even when a rebuild happens during it.
• Full suite on pyglet 3.0.dev11: 1485 passed. The 3 failures are the render tests that only fail on my machine.
• Changelog: added as one more item in the unreleased create_text_sprite fixes entry.

Once this is merged, #2917 can be updated from development and its CI should pass.

🤖 Generated with Claude Code