fix(gpu): cap GPU readback staging buffers to prevent memory exhaustion

[mjohnson7cd]

Summary

Fixes storytold/artcraft#2081 - Rendering stops at frame 970/1570, PC lags until app closed on high-end system (i7-12700KF, RTX 3090, 32GB RAM).

Problem

GPU readback staging buffers accumulated unbounded when layer bounds changed every frame (animated scale, blur radius, growing shapes). Each 1080p RGBA f32 buffer ≈ 33 MB in system RAM. At 970 frames × 33 MB ≈ 32 GB, the system exhausted available memory.

The buffers are kept for reuse (up to 2 per size), but there was no limit on total memory across all sizes.

Solution

Capped GPU readback staging buffers to 256 MB total. When adding a new buffer would exceed the cap, drop buffers of other sizes while keeping the current size for reuse. Renders with consistent frame sizes still benefit from buffer reuse.

Implementation

Changed: crates/gpu/src/context.rs, crates/gpu/tests.rs (+34/-1)

• Added 256 MB budget for readback staging buffers
• Eviction policy: drop other sizes when budget exceeded, keep current size
• New test: readback_staging_stays_bounded_across_sizes

Test Results

Without fix: Test reads 120 different sizes (~480 MB), holds 620 MB, takes 28.5s (Apple M4/Metal)
With fix: Same test passes in 0.54s, stays within 256 MB budget

✓ cargo fmt
✓ cargo clippy --workspace --all-targets -- -D warnings
✓ cargo test -p effectcraft-gpu (204 tests passed)

Notes

• Verified on Metal backend; not tested on Windows/DX12/Vulkan with RTX 3090
• Would benefit from reporter's project file to confirm exact reproduction
• Small rebase conflict possible with open PRs GPU: resident effect stacks, in-frame texture reuse, upload-cache budget, stable footprint keys (#687) #688, Cache/GPU: adjustment footprint ignores the effect stack; upload starvation is reported (#687) #701 (also touch GPU cache code)

---

— sent on behalf of @mjohnson7cd by Claude Code