(feat): Avoid GPU realloc on CuVector shrink

[mgyoo86]

Background

Pool operations frequently resize! backing vectors. On CPU, shrinking preserves capacity — no problem. On CUDA, shrinking below 25% of capacity triggers a full GPU reallocation (alloc → copy → free). resize!(v, 0) in particular always reallocates, defeating the pool's zero-allocation guarantee.

Solution

_resize_without_shrink! changes only the logical size (dims) on shrink, preserving the GPU allocation. When the vector grows back to the same size, the existing memory is reused with no reallocation.

Impact

• Acquire path: reusing a slot for a smaller array no longer triggers GPU realloc
• Safety Level 1 parity: released slots can now be structurally invalidated to length 0 (poison → shrink to 0), matching CPU's resize!(vec, 0) behavior. Previously CUDA could only poison without structural invalidation due to the memory-free issue
• CPU–CUDA safety symmetry: Levels 0–3 now check identically across both backends

Compatibility

• Relies on CuArray internals (:dims, .maxsize) — @assert hasfield compile-time guard added (tested with CUDA.jl v5.x)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 74.57%. Comparing base (7dc30c8) to head (29c110b).
⚠️ Report is 1 commits behind head on master.

❗ There is a different number of reports uploaded between BASE (7dc30c8) and HEAD (29c110b). Click for more details.

HEAD has 3 uploads less than BASE

Flag	BASE (7dc30c8)	HEAD (29c110b)
	6	3

Additional details and impacted files

@@             Coverage Diff             @@
##           master      #28       +/-   ##
===========================================
- Coverage   96.79%   74.57%   -22.22%     
===========================================
  Files          14       14               
  Lines        2620     2616        -4     
===========================================
- Hits         2536     1951      -585     
- Misses         84      665      +581     

see 7 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

Adds a CUDA-specific resize strategy to avoid costly GPU reallocations when pooled CuVectors are shrunk (especially to length 0), restoring the pool’s “no GPU allocation on reuse” behavior and making CUDA safety level 1 structurally invalidate released slots like the CPU backend.

Changes:

• Introduces _resize_without_shrink! for CuVector shrink operations (dims-only shrink; grow delegates to resize!) and uses it in the CUDA acquire path.
• Updates CUDA safety level 1 invalidation to shrink released vectors to length 0 after poisoning (without freeing GPU memory).
• Adjusts CUDA tests to assert length→0 after rewind and adds pointer-preservation tests for _resize_without_shrink!.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
ext/AdaptiveArrayPoolsCUDAExt/acquire.jl	Adds _resize_without_shrink! and uses it to avoid GPU realloc on shrink in get_view!.
ext/AdaptiveArrayPoolsCUDAExt/debug.jl	Level 1 invalidation now shrinks released vectors to length 0 using _resize_without_shrink! after poisoning.
test/cuda/test_cuda_safety.jl	Updates safety tests to reflect structural invalidation (length 0) and verifies poisoning via re-acquire.
test/cuda/test_allocation.jl	Adds tests ensuring shrink/grow-back preserves the GPU pointer when within maxsize.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.