Guard the ReducedTensor allocation readings on coverage

CHANGELOG.md

@@ -12,6 +12,16 @@ tags. It is named as a gap rather than reconstructed, because a changelog assemb
 fact loses exactly the reasoning that makes it worth keeping. Every release from `v0.2.0` on
 is written up below, newest first.
 
+## [Unreleased] — targeting 0.4.1
+
+### Changed
+
+- **The allocation readings of `ReducedTensor` are taken only where the runner does not instrument
+  coverage**, as the other `== 0` readings of the suite are. The `== 0` readings of `getindex` and
+  `copyto!` and their `> 0` control in `test/reduced_tensor.jl` sit behind
+  `Base.JLOptions().code_coverage == 0`, because coverage instrumentation allocates on its own
+  account. Tests only; nothing in the package changes.
+
 ## [0.4.0] — 2026-10-30
 
 ### Breaking Changes

test/reduced_tensor.jl

@@ -225,9 +225,14 @@ allocations(f::F, a::A, b::B) where {F, A, B} = (f(a, b); @allocated f(a, b))
             local gt = neighbourhood_tensor(Xoshiro(1), DT, g, w)
             local rt = ReducedTensor(gt, rand(Xoshiro(2), DT, n, 3), rand(Xoshiro(3), DT, n, 2))
             local out = Array{DT}(undef, size(rt))
-            @test allocations((r, o) -> r[2, 1, 7], rt, out) == 0
-            @test allocations((r, o) -> copyto!(o, r), rt, out) == 0
-            @test allocations((r, o) -> collect(r), rt, out) > 0
+
+            # Coverage instrumentation allocates on its own account, so the readings are taken
+            # only where the runner does not instrument the code.
+            if Base.JLOptions().code_coverage == 0
+                @test allocations((r, o) -> r[2, 1, 7], rt, out) == 0
+                @test allocations((r, o) -> copyto!(o, r), rt, out) == 0
+                @test allocations((r, o) -> collect(r), rt, out) > 0
+            end
         end
     end