[MatejKozlovsky]

Fixes #687.

The overall series vs this PR. This is PR 2 of a series that reworks GPU export. The full series,
measured on a test PC (Ryzen 7 5800X, RTX 3090 24 GB, Windows 10) with the same 10 s 4K comp: 0.7.0
stock --gpu 575.1 s vs the full series 33.5 s (~17.2×). This PR alone delivers much less; its
numbers are below. The batch-parallel export that keeps frames off the GPU on main is addressed by a
later PR (the scheduler); here exports are measured with RAYON_NUM_THREADS=1 so frames stay on the GPU.

Changes (4 commits, one change each)

1. Render: an adjustment layer's footprint key ignores its effects (the footprint is the source
   through the masks), so a time-dependent effect such as Noise no longer re-keys it every frame.
2. GPU: an Enc reuses the working textures it released at once. Its later commands run after
   everything it recorded before (one queue, recording order), so it may overwrite them; other encoders
   only get them after the Enc is gone, through the existing pool.
3. GPU: GpuContext::set_upload_budget / upload_budget (default unchanged, 1536 MB), and one
   warning when the upload cache is full and starts evicting layer textures.
4. GPU: walk::draw_resident_fx: a plain layer's effect stack runs resident on the GPU (GpuFx, in
   effect space, as 0.7.0 defines it, then rebased to layer space) and is placed directly; CPU-only effects
   in the stack read back, run and upload as before. A layer whose processed buffer is cached and unchanged
   since the last frame (Renderer::layer_is_static) keeps the usual path, so it uploads once instead of
   re-running its effects every frame. Layers with layer styles, Preserve Transparency or a track matte
   keep the usual path.

Numbers (measured, test PC, main 2ec87e9 vs main + this PR)

4K comp with effect stacks, 3 s / 90 frames, --gpu, RAYON_NUM_THREADS=1:

	wall	fps	GPU util avg/max	VRAM peak
main, CPU (reference)	173.3 s	0.52	–	–
main --gpu	81.1 s	1.11	29% / 100%	17,185 MiB
main + this PR --gpu	49.3 s	1.83 (1.65×)	18% / 95%	9,376 MiB

Both GPU exports differ from the CPU export (MD5), so the frames really came from the GPU. They are
bit-identical to each other (PSNR ∞) and both 50.9 dB average PSNR vs CPU (Y 49.7 dB). No GPU-declined or
upload-cache-full lines in either log.

effectcraft-cli bench --gpu, same comp, median of 5 (ms/frame):

	main	this PR
4K GPU cold	1181.0	491.2
4K GPU warm	208.1	128.4
4K GPU view	172.4	64.1
1080p GPU cold	325.6	177.9
1080p GPU warm	51.3	48.9
1080p GPU view	35.4	22.1

GPU vs CPU: 0.000% of pixels over 1/255 on both. Upload traffic on this comp is 0.0 MB/frame warm on both
(nothing in it re-keys), so the footprint-key and upload-budget changes don't show here; their tests
measure them. An earlier version of this PR re-ran static effect stacks every frame and was slower at
1080p warm (83.7 ms vs 51.8 ms); commit 4's layer_is_static check fixed that (48.9 ms above).

Tests

• effect_stacks_render_resident: GPU vs CPU parity, 8 and 32 bpc, full and half resolution: GPU effects
  with a CPU-only effect in the middle, opacity, a blend mode, rotation, and a shape layer (effect space
  from its bounds' corner).
• static_effect_layers_draw_from_the_cache: a static layer with effects uploads nothing after it is
  cached, frames repeat exactly, matches the CPU.
• texture_reuse_does_not_change_output: a busy frame is bit-identical with in-Enc reuse on and off,
  and matches the CPU. enc_reuses_its_own_released_textures: the Enc gets its own texture back at once;
  another Enc does not.
• adjustment_footprint_uploads_once_under_time_dependent_effects: the footprint key is equal at two
  times (no GPU needed); on the GPU, frames 3 and 4 upload 0 bytes (before the fix: 64,000 bytes, the
  footprint, every frame); output still changes every frame and matches the CPU.
• upload_budget_bounds_the_layer_textures_kept: within the budget static layers upload once; under a
  1 MB budget they upload again; the frame is the same.
• All existing GPU parity suites now go through the resident path and pass.

CI

Rust 1.95, on top of main 2ec87e9:

• cargo fmt --check: clean. xtask layers, assets (39 assets), wasm (32 crates): ok.
• Workspace tests: all pass except 2 that also fail on unmodified main (same pixels): effectcraft-gpu
tests_fx_pixel2::stylize (cchextile, 2 of 770 px) and
  tests_fx_stylize::posterize_thresholds_strobe_vignette (ccthresholdrgb, 13 of 5917 px), on a software
  Vulkan adapter.
• clippy -D warnings: clean for this PR's code. Upstream errors on unmodified main, not touched here:
  collapsible_match in crates/psd/src/lib.rs:660, crates/text/src/path_text.rs:36,
  crates/effects/src/audio_fx.rs:249, crates/effects/src/generate3.rs:507, and nonminimal_bool in
  crates/ui-egui/src/panels/text_panels.rs:856.
• CPU export (no --gpu) of the 4K comp is byte-identical between main and this PR.

Limitations

• Exports on main still render a batch of frames in parallel, which keeps many frames off the GPU; that
  is the next PR (the scheduler). The export numbers above use RAYON_NUM_THREADS=1 for that reason.
• Layer styles, Preserve Transparency and track mattes keep the existing path for effect layers.
• The upload budget is an API; picking it from the device's memory is left to the embedder (a later PR).

[MatejKozlovsky]

Merged current main into this branch to resolve the conflict; build and tests pass locally. #688 and #898 are both clean now and independent. If it helps, #688 then #898 is a sensible merge order.

[echelon]

Merged. Thank you for the contribution, @MatejKozlovsky! It landed on main together with a batch of other community PRs, after a full workspace build and test run. If you are not on our Discord yet, please join us: https://discord.gg/artcraft