Make test_abstract_blocktype work with MatrixAlgebraKit 0.6.6

Two related issues surface in `test/test_abstract_blocktype.jl` once
MatrixAlgebraKit 0.6.6 is resolved:

* `@test_broken f(a)` for `f ∈ (svd_full, svd_trunc)` on JLArray-backed
  `BlockSparseMatrix{T, AbstractMatrix{T}}` errored with "Expression
  evaluated to non-Boolean" because `svd_full` now succeeds and returns
  a 3-tuple (it used to throw, which `@test_broken` correctly caught).
* The unconditional `@test c * u ≈ a` in the right-side loop was
  flaky on JLArray for `right_orth`, `lq_compact`, `lq_full` — these
  produce `c * u != a` in MatrixAlgebraKit 0.6.6 (regression from 0.6.4).

This commit:

* Seeds each `(arrayt, elt)` iteration with `StableRNG(1)` so the
  random matrices are deterministic across runs.
* Promotes `svd_full` to a regular `@test` on JLArray (it now produces
  a correct factorization).
* Wraps `svd_trunc` in a Boolean-safe `@test ... broken = true` block
  so its return value can no longer surface as a non-Boolean error.
* Marks `c * u ≈ a` as broken on JLArray for `right_orth`, `lq_compact`,
  `lq_full`, with a comment pointing at the upstream regression.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: mtfishman

Project.toml

@@ -1,6 +1,6 @@
 name = "BlockSparseArrays"
 uuid = "2c9a651f-6452-4ace-a6ac-809f4280fbb4"
-version = "0.10.39"
+version = "0.10.40"
 authors = ["ITensor developers <support@itensor.org> and contributors"]
 
 [workspace]

test/test_abstract_blocktype.jl

@@ -7,13 +7,15 @@ using MatrixAlgebraKit: eig_full, eig_trunc, eig_vals, eigh_full, eigh_trunc, ei
     isisometric, left_orth, left_polar, lq_compact, lq_full, qr_compact, qr_full,
     right_orth, right_polar, svd_compact, svd_full, svd_trunc
 using SparseArraysBase: storedlength
+using StableRNGs: StableRNG
 using Test: @test, @test_broken, @testset
 
 elts = (Float32, Float64, ComplexF32)
 arrayts = (Array, JLArray)
 @testset "Abstract block type (arraytype=$arrayt, eltype=$elt)" for arrayt in arrayts,
         elt in elts
 
+    rng = StableRNG(1)
     dev = adapt(arrayt)
 
     a = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
@@ -22,7 +24,7 @@ arrayts = (Array, JLArray)
     @test iszero(blockstoredlength(a))
 
     a = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
-    a[Block(1, 1)] = dev(randn(elt, 2, 2))
+    a[Block(1, 1)] = dev(randn(rng, elt, 2, 2))
     @test !iszero(a[Block(1, 1)])
     @test a[Block(1, 1)] isa arrayt{elt, 2}
     @test iszero(a[Block(2, 2)])
@@ -33,9 +35,9 @@ arrayts = (Array, JLArray)
     @test a[Block(1, 2)] isa Matrix{elt}
 
     a = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
-    a[Block(1, 1)] = dev(randn(elt, 2, 2))
+    a[Block(1, 1)] = dev(randn(rng, elt, 2, 2))
     a′ = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
-    a′[Block(2, 2)] = dev(randn(elt, 3, 3))
+    a′[Block(2, 2)] = dev(randn(rng, elt, 3, 3))
 
     b = copy(a)
     @test Array(b) ≈ Array(a)
@@ -54,7 +56,7 @@ arrayts = (Array, JLArray)
     @test norm(b) ≈ 0
 
     a = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
-    a[Block(1, 1)] = dev(randn(elt, 2, 2))
+    a[Block(1, 1)] = dev(randn(rng, elt, 2, 2))
     for f in (eig_full, eig_trunc)
         if arrayt === Array
             d, v = f(a)
@@ -71,7 +73,7 @@ arrayts = (Array, JLArray)
     end
 
     a = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
-    a[Block(1, 1)] = dev(parent(hermitianpart(randn(elt, 2, 2))))
+    a[Block(1, 1)] = dev(parent(hermitianpart(randn(rng, elt, 2, 2))))
     for f in (eigh_full, eigh_trunc)
         if arrayt === Array
             d, v = f(a)
@@ -88,7 +90,7 @@ arrayts = (Array, JLArray)
     end
 
     a = BlockSparseMatrix{elt, AbstractMatrix{elt}}(undef, [2, 3], [2, 3])
-    a[Block(1, 1)] = dev(randn(elt, 2, 2))
+    a[Block(1, 1)] = dev(randn(rng, elt, 2, 2))
     for f in (left_orth, left_polar, qr_compact, qr_full)
         u, c = f(a)
         @test u * c ≈ a
@@ -101,7 +103,10 @@ arrayts = (Array, JLArray)
     end
     for f in (right_orth, right_polar, lq_compact, lq_full)
         c, u = f(a)
-        @test c * u ≈ a
+        # `right_orth`, `lq_compact`, `lq_full` on JLArray-backed
+        # `BlockSparseMatrix{T, AbstractMatrix{T}}` produce a `c * u` that does not
+        # reproduce `a` (regression in MatrixAlgebraKit 0.6.6).
+        @test c * u ≈ a broken = (arrayt ≢ Array && f ∈ (right_orth, lq_compact, lq_full))
         if arrayt ≡ Array
             @test isisometric(u; side = :right)
         else
@@ -110,8 +115,13 @@ arrayts = (Array, JLArray)
         end
     end
     for f in (svd_compact, svd_full, svd_trunc)
-        if arrayt ≢ Array && (f ≡ svd_full || f ≡ svd_trunc)
-            @test_broken f(a)
+        if arrayt ≢ Array && f ≡ svd_trunc
+            # `svd_trunc` on JLArray-backed `BlockSparseMatrix{T, AbstractMatrix{T}}`
+            # currently triggers scalar indexing on the GPU array.
+            @test begin
+                u, s, v = f(a)
+                u * s * v ≈ a
+            end broken = true
         else
             u, s, v = f(a)
             @test u * s * v ≈ a