[AntonOresten]

Lengths are checked. Every length the left side of a pattern fixes is now compared with the input before anything reshapes it. That covers integers, 1, and names given a length by keyword. Before, "4 b -> b" reduced an axis of length 3, and a=4 for an ungrouped a was never looked at, because a known length only entered the plan inside a group, as a split. A group with one name left to infer must be divisible by the rest. Errors name the input axis, also past an ellipsis:

julia> @reduce(sum, rand(3, 2), "4 b -> b")
ERROR: DimensionMismatch: Input axis 1 has length 3, but the pattern expects 4

"1 b -> b" on a length-3 axis still throws. That is the same rule, now with the same message.

1 inside parentheses is a unit, as in Python einops: (b a 1) is (b a), and () or (1) is a length-one axis. An entry the plan builders do not recognize now raises instead of being skipped, which is what turned (b a 1) into an empty reshape (#60). Other integers say where they are allowed: reduced on the left of reduce, or created on the right of repeat.

$k interpolates a length (#35): an axis of length k, as if the value were written there, the way Python einops users write f"({k} a)". Like an integer other than 1, it can only be reduced or created, and each occurrence is an axis of its own:

@reduce(mean, x, "(h $k) (w $k) c -> h w c")   # k×k pooling
@repeat(x, "L1 L2 -> ($k L1) ($k L2)")         # #36's use case

The value never enters the pattern's type (#34):

• The shorthand macros pass it on as a hidden keyword.
• einops"..." returns a new public BoundPattern, which shows as the pattern with the values in place.

A length computed at run time therefore leaves the call type stable, and @inferred tests cover this. An integer built into a --> tuple pattern still cannot do that, since the type can't tell 1 from 3; the --> docstring now says so and points to the alternatives. For the same reason a value of 1 stays an axis of length one rather than becoming a unit: reduced or created, that gives the same result, but $k is rejected in rearrange whatever its value.

A name is an index, so duplicate names stay rejected. #36 asked for duplicates as a size hint, but an index twice on one side means a diagonal, which is einsum's business ("k k ->" is the trace, there and in Python). $k covers the size-hint use case instead. Keeping an axis through rearrange takes a name and a keyword, "(h d) n -> d n h"; h. einsum rejects $: such an axis could only be summed out, which a name does too, and einsum isn't type stable to begin with.

Version 0.3.0 requires Julia 1.10 and exports @rearrange, @reshape, @reduce, @repeat and @einsum, which also makes the README quickstart work after using Einops. Breaking for:

• calls that passed a wrong length, which used to give a wrong result silently;
• code loading TensorCast.jl (which exports @reduce) or Einsum.jl (which exports @einsum) alongside Einops; those names now need qualifying.

Tests: test/axis_lengths.jl is new. The full suite passes on Julia 1.10 and 1.13. Zygote passes on 1.12 and Reactant on 1.13, each with new tests for checked and interpolated lengths, including a wrong length throwing while tracing. Doctests and the docs build are clean.

Closes #34, closes #35, closes #36, closes #60.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 99.32432% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 99.49%. Comparing base (ff0f1e8) to head (7eff20c).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/repeat.jl	87.50%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #62      +/-   ##
==========================================
- Coverage   99.56%   99.49%   -0.07%     
==========================================
  Files          15       15              
  Lines         683      793     +110     
==========================================
+ Hits          680      789     +109     
- Misses          3        4       +1     

☔ 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.