Fix ArrayPartition A[:] DimensionMismatch (#661) - #664

Summary

A[:] on ArrayPartition was dispatching to the (colon, j...) method introduced in PR #239 with an empty j, which broadcast getindex over zero indices and threw DimensionMismatch. The fix requires at least one index after Colon (j, js...) so A[:] uses the existing flattening ::Colon method, while A[:, j...] from #239 is unchanged.

Bisect: first bad commit is ba36f16 ("Fix (colon,j...) indexing of ArrayPartition", PR #239). Parent ba36f16^ returns the flattened vector; ba36f16 breaks A[:] (initially MethodError on isless(::Int, ::Colon)); 1ce74b8b ("oops add the dispatch") changes the failure to the current DimensionMismatch.

Failing before (src fix stashed; regression testset)

DimensionMismatch: arrays could not be broadcast to a common size: a has axes Base.OneTo(2) and b has axes Base.OneTo(0)
Test Summary:                               | Pass  Error  Total  Time
ArrayPartition A[:] flattening (issue #661) |    4      9     13  5.2s

(The 4 passes are range indexing and (colon, j...) checks; all A[:] asserts error.)

Passing after

Test Summary:                               | Pass  Total  Time
ArrayPartition A[:] flattening (issue #661) |   13     13  2.3s

Test group run

GROUP=Core julia --project=. -e 'using Pkg; Pkg.test()' — passed.

Test Summary:    | Pass  Total   Time
Partitions Tests |  128    128  38.4s
...
Measurement Tests |    2      2  0.1s
     Testing RecursiveArrayTools tests passed

What was not verified

• Downstream / AD / GPU / QA / SymbolicIndexingInterface groups
• Behavior of setindex! with A[:] (out of scope)
• Older Julia versions beyond 1.12.4 used here

What a reviewer should push back on

• Whether requiring j, js... (at least one index) is the preferred fix vs making the flatten ::Colon method win on specificity another way
• Whether nested ArrayPartition flatten via Chain is the desired semantics for A[:] in all cases
• Docstring wording mentioning issue ArrayPartition A[:] throws DimensionMismatch #661

Fixes #661

Please ignore this draft until reviewed by @ChrisRackauckas.

Risk assessment

• Risk: low
• Blast radius: a single getindex(::ArrayPartition, ::Colon, j, js...) signature, tightened so that A[:] dispatches to the existing flattening getindex(::ArrayPartition, ::Colon) method. A[:, j...] behaviour is unchanged. A[:] threw on every ArrayPartition since Fix (colon,j...) indexing of ArrayPartition #239 and now returns the flattened vector. No other public API change.
• Evidence: bisected to ba36f16 (Fix (colon,j...) indexing of ArrayPartition #239). The new tests fail on master with DimensionMismatch and pass on the PR head. I reran this independently: PR head 5/5 pass, master src gives the 2 DimensionMismatch errors.
• Independent review: Claude Code (Opus 5.5, head session; author is Cursor Auto) rated it low, high confidence, verdict MERGE. One docstring cleanup was requested (drop the issue reference).
• Merge: auto-merge candidate once CI is green.

🤖 Generated with Cursor Agent CLI 2026.09.26-dd393fe (model: auto), transcript /home/crackauc/sandbox/goals/issue-backlog/jobs/fix-RecursiveArrayTools.jl-661/log.txt on amdci2.julia.csail.mit.edu

Made with Cursor

Risk assessment and review by 🤖 Claude Code (model: claude-opus-5-5[1m]), https://claude.ai/code/session_01NdFVhANTF4nepTbqYxjfRP