[devmotion]

unzip(xs::Tuple) can't be inferred: ntuple(i -> map(Get(i), xs), length(x1)) builds Get(i) from a runtime i, so the components are inferred with Any elements:

julia> Base.infer_return_type(ChainRules.unzip, Tuple{NTuple{3,Tuple{Float64,Float64}}})
Tuple{Tuple{Any, Any, Any}, Tuple{Any, Any, Any}}

This becomes visible with JuliaDiff/ChainRulesCore.jl#724, where the cotangent of a tuple-valued broadcast stays a Tuple instead of being collected into a Vector. unzip_broadcast then takes its tuple branch, and test_rrule(copy∘broadcasted, BT1, atan, Tuple(rand(3)), Tuple(rand(3))) fails its inference check (the only failure in the ChainRules integration job of that PR).

This PR dispatches on Tuple{Vararg{NTuple{N,Any}}}, a non-empty tuple of tuples of equal length N, and forwards to the existing unzip(xs, Val(N)) method, as the array path already does. After this change:

julia> Base.infer_return_type(ChainRules.unzip, Tuple{NTuple{3,Tuple{Float64,Float64}}})
Tuple{Tuple{Float64, Float64, Float64}, Tuple{Float64, Float64, Float64}}

Invalid inputs now throw descriptive errors instead of failing in first, throwing a BoundsError, or silently dropping elements:

input	before	after
()	ArgumentError: tuple must be non-empty	ArgumentError: unzip cannot determine the number of components of an empty tuple
((1, 2), 3)	BoundsError	ArgumentError: unzip only accepts arrays or tuples of tuples
((1, 2), [3, 4])	((1, 3), (2, 4))	ArgumentError: unzip only accepts arrays or tuples of tuples
((1, 2), (3,))	BoundsError	DimensionMismatch: unzip requires tuples of equal length, got lengths (2, 1)
((1,), (2, 3))	((1, 2),)	DimensionMismatch: unzip requires tuples of equal length, got lengths (1, 2)

Inside ChainRules, all elements come from the same pullback, so they are always tuples of the same length. The empty-tuple broadcast test (atan over () and ()) does not reach unzip.

This supersedes the unzip part of #858. The _bc_cotangent part of #858 is no longer needed with JuliaDiff/ChainRulesCore.jl#724. It is independent of #857; both bump the version to 1.74.9, so whichever is merged second needs a trivial rebase.

Tests

@inferred tests for valid inputs, and tests for each of the invalid inputs above. Locally on Julia 1.13:

• with the final code, test/unzipped.jl passes with Throw when broadcasting over structural tangents ChainRulesCore.jl#724;
• with an earlier version of this PR, which only replaced the ntuple call with unzip(xs, Val(length(x1))) and gives the same results for valid inputs, test/unzipped.jl and rulesets/Base/broadcast passed with ChainRulesCore 1.26.7, and the failing atan test passed with Throw when broadcasting over structural tangents ChainRulesCore.jl#724.

The full test suite was not run locally.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.63%. Comparing base (7fb6cb1) to head (0ba0b12).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #859   +/-   ##
=======================================
  Coverage   96.63%   96.63%           
=======================================
  Files          28       28           
  Lines        3596     3598    +2     
=======================================
+ Hits         3475     3477    +2     
  Misses        121      121           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

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