(fix): Use legacy N-way cache path on Julia 1.11

[mgyoo86]

Problem

On Julia 1.11, length(::Array) is computed from the backing Memory object's length (arraylen builtin), not from prod(size(arr)). This was fixed in Julia 1.12.

The setfield!(:ref)-based wrapper reuse in get_array! copies the entire MemoryRef from the backing vector. Since pool slots grow but never shrink, the backing vector can be larger than prod(dims):

# Julia 1.11
v = acquire!(pool, Float64, 10, 10)   # backing vec sized to 100
rewind!(pool, Float64)
b = acquire!(pool, Float64, 5, 4)     # same slot, backing vec still 100
length(b)   # 100 ← WRONG (should be 20)
vec(b)      # size (100,) ← breaks downstream code

This caused DimensionMismatch errors in production.

Fix

Move the modern/legacy version gate from v"1.11-" to v"1.12-":

• Julia ≤1.11: Uses the legacy N-way set-associative cache path (unsafe_wrap + cache). Zero-alloc on cache hit (up to CACHE_WAYS patterns per slot, default 4).
• Julia 1.12+: Uses the modern setfield!(:ref, :size) path. Zero-alloc unconditionally (unlimited dimension patterns).

GPU extensions (CUDA, Metal) are also gated at v"1.12-" because they import modern-path-only functions (_store_arr_wrapper!) from the base package.

Impact

Julia version	CPU path	GPU pooling	Zero-alloc?
≤1.10	legacy (N-way cache)	disabled	cache hit: yes
1.11	legacy (N-way cache)	disabled	cache hit: yes
1.12+	modern (setfield!)	enabled	always

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 95.48%. Comparing base (3118fd0) to head (a2f444c).
⚠️ Report is 2 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #41      +/-   ##
==========================================
- Coverage   96.03%   95.48%   -0.56%     
==========================================
  Files          14       14              
  Lines        3232     3232              
==========================================
- Hits         3104     3086      -18     
- Misses        128      146      +18     

Files with missing lines	Coverage Δ
src/AdaptiveArrayPools.jl	100.00% <ø> (ø)
src/acquire.jl	90.67% <ø> (-5.70%)	⬇️
src/bitarray.jl	88.88% <ø> (-2.23%)	⬇️
src/convenience.jl	98.70% <ø> (ø)
src/debug.jl	95.00% <ø> (ø)
src/legacy/state.jl	96.27% <ø> (ø)
src/types.jl	85.52% <ø> (-3.95%)	⬇️

... and 1 file 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

Adjusts version gating so Julia 1.11 uses the legacy N-way cache path (avoiding incorrect length(::Array)/vec behavior), while keeping the modern setfield! wrapper-reuse path and GPU extensions enabled only on Julia 1.12+.

Changes:

• Moved the modern/legacy version gate from v"1.11-" to v"1.12-" in the package entrypoint and related code paths.
• Updated tests to only exercise setfield!-based invalidation/zero-allocation behavior on Julia 1.12+ and use legacy expectations on ≤1.11.
• Gated CUDA/Metal extensions and their tests to Julia 1.12+ and updated related documentation/comments.

Reviewed changes

Copilot reviewed 19 out of 19 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/AdaptiveArrayPools.jl	Switches the include/version gate to VERSION >= v"1.12-" so Julia 1.11 loads legacy implementation.
src/acquire.jl	Updates documentation/comments for setfield!-based reshape/acquire path to be 1.12+.
src/bitarray.jl	Updates header/comments to reflect BitArray setfield-based reuse is for Julia 1.12+.
src/convenience.jl	Updates reshape! doc wording to 1.12+ path.
src/debug.jl	Gates wrapper-mutation checks to 1.12+ and updates related assumptions/comments.
src/types.jl	Updates type docs to describe arr_wrappers cache as 1.12+ feature.
src/legacy/state.jl	Updates legacy header/docs to indicate legacy applies through Julia ≤1.11.
ext/AdaptiveArrayPoolsCUDAExt/AdaptiveArrayPoolsCUDAExt.jl	Requires Julia 1.12+ for CUDA pooling and updates warning text/comments.
ext/AdaptiveArrayPoolsCUDAExt/types.jl	Updates CUDA pool type docs to match Julia 1.12+ CPU design reference.
ext/AdaptiveArrayPoolsCUDAExt/acquire.jl	Updates comments to reference CPU’s Julia 1.12+ approach.
ext/AdaptiveArrayPoolsMetalExt/AdaptiveArrayPoolsMetalExt.jl	Requires Julia 1.12+ for Metal pooling and updates warning text/comments.
ext/AdaptiveArrayPoolsMetalExt/types.jl	Updates Metal pool type docs to match Julia 1.12+ CPU design reference.
test/runtests.jl	Updates version-based test selection and helpers to gate modern-path tests on 1.12+.
test/test_zero_allocation.jl	Updates per-iteration allocation tolerance comment/version gate to 1.12+.
test/test_safety.jl	Gates setfield!-based invalidation assertions to run only on Julia 1.12+.
test/test_reshape.jl	Gates “zero allocation reshape” tests to Julia 1.12+.
test/test_nway_cache.jl	Updates comments describing unlimited-pattern zero-alloc behavior to Julia 1.12+.
test/cuda/runtests.jl	Skips CUDA extension tests on Julia < 1.12 and updates messaging.
test/metal/runtests.jl	Skips Metal extension tests on Julia < 1.12 and updates messaging.

Comments suppressed due to low confidence (1)

src/convenience.jl:352

• This docstring still says “On Julia 1.10 and CUDA, falls back to Base.reshape”, but after moving the version gate to 1.12 the fallback behavior also applies on Julia 1.11 (since the setfield!-based reshape implementation isn’t included there). Please update the version wording here so it matches the new gating.

On Julia 1.12+:
- If `ndims(A) == length(dims)` (same dimensionality), `reshape!` mutates `A`
  in-place by changing its size. This differs from `Base.reshape`, which always
  returns a new wrapper.
- For cross-dimensional reshapes (`ndims(A) != length(dims)`), the returned
  `Array` wrapper is taken from the pool's internal cache and may be reused
  after `rewind!` or pool scope exit.

As with all pool-backed objects, the reshaped result must not escape the
surrounding `@with_pool` scope.

On Julia 1.10 and CUDA, falls back to `Base.reshape`.

---

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