Textures staged with Stage::set_images after a render are silently not applied

[schell]

Textures staged with Stage::set_images after a render are silently not applied

Repro

1. Create a Stage, primitives, and a Material; attach the material and render once (any frame).
2. Call Stage::set_images([...]) to stage textures in the atlas.
3. material.set_albedo_texture(&entries[0]) and render again.

Result: the texture is silently not sampled — the material falls back to the
albedo factor. Renders for different textures are byte-identical (0 pixels differ).

Isolated trigger

Rendering before set_images is the trigger. Creating the material (and
attaching it to primitives) before set_images is harmless — the bug only
appears when a render happens between stage creation and set_images. Calling
set_images before the first render works fine.

Experiment (two cubes sharing one material, 512×512):

order	result
render → set_images → texture → render	flat white cubes; renders for sandstone/dirt byte-identical (6,329 bytes both)
set_images → render → texture → render	textures render: 193KB/159KB PNGs, 19,077 unique colors vs 23 for the flat render

Suspected area

Stage::commit (crates/renderling/src/stage/cpu.rs, the
materials_atlas_texture_was_recreated → primitive_bind_group.invalidate()
path) apparently does not fully propagate once a render has already cached the
renderlet bindgroup — the atlas texture view and/or materials slab buffer bound
at first render persist, so the shader's texture lookup fails and silently
falls back to the albedo factor.

Suggested fix / acceptance

• Minimal test in the renderling crate: render → set_images → set texture →
  render, assert the texture is visible (unique-color count or img_diff)
• Ideally also make the silent fallback loud (log/assert) when a material
  references a texture that can't be sampled

Context

Found while writing the materials manual chapter (crates/examples/src/material.rs
in the docs-missing-manual-buildout worktree; the example currently works
around it by staging all images up front, which the set_images docs recommend
anyway).