Restore inferrable return type for block tensor slicing

[lkdvos]

Fixes a type-inference regression introduced in v0.3.16 by #72, and bumps the version to v0.3.17.

The regression

_slice_getindex chose between returning a single block and a sliced block tensor with a runtime inds isa NTuple{M, Int} check, and _slice_getindex_single reached that check with an index tuple built as ntuple(i -> i == d ? only(indices) : 1, numind(t)) — element type Union{Int, UnitRange{Int}}, so the check could not be resolved statically and the union leaked into the return type.

On an MPS-environment-shaped tensor (size (1, 3, 1)):

	t[2:(end - 1)] inferred as
v0.3.15	BlockTensorMap{…}
v0.3.16	Union{BlockTensorMap{…}, TensorMap{…}}

The same applied to t[:], t[1:2:3] and t[[1, 3]]; a logical mask (t[[true, false, true]]) inferred as Any, for a second reason — a mask may consume several dimensions, so Base.to_indices cannot infer one mixed with the 1 fillers.

The fix

Make the single-block-vs-slice choice by dispatch (_slice_getindex_full), and fill the trivial dimensions with 1:1 rather than 1 so Int stays out of the index tuple. Masks are normalized to an index vector before being placed.

After the fix, every index form on every shape infers concretely — Int, UnitRange, StepRange, Colon, index vector and logical mask, single-index and full-rank, dense and sparse.

Why it matters

MPSKit's Jordan-form derivative operators store O3 = typeof(A) where A is built from GL[2:(end - 1)], so the widened slice type made AC_hamiltonian infer as the bare abstract JordanMPO_AC_Hamiltonian, and from there excitations with ChepigaAnsatz/ChepigaAnsatz2 inferred Tuple{Any, Vector} — a CI failure in MPSKit's @testinferred tests. There is no measurable runtime cost (DMRG, 20 sites, χ=32, 5 sweeps: 0.353 s on v0.3.15 vs 0.363 s on v0.3.16), so this is purely a type-stability contract regression.

Verification

• Full test suite green on Julia 1.12.6, Aqua included (no new method ambiguities).
• Slicing produces identical types and values to released v0.3.16 across a battery of 13 index forms × 4 tensor shapes (dense/sparse, one- and three-dimensional). The remaining diffs against v0.3.15 are all v0.3.16's intentional fixes (repeated indices, step ranges, uninitialized blocks).
• With the fix dev'd in, MPSKit's failing excitations "finite" testset passes 25/25 (it was 21 pass / 4 fail), on both Julia 1.10.11 and 1.12.6.

Tests

test/abstracttensor/indexing.jl gains:

• a single-index slicing testset on a multi-index tensor with one nontrivial dimension. The existing single-index tests use a genuinely one-dimensional tensor, which takes the M == numind(t) fast path, so this path had no coverage at all.
• an indexing inference sweep asserting a concrete return type for every index form across six tensor shapes, so a future rewrite of the index handling is caught for all of them rather than the hand-picked cases.

Both fail on v0.3.16 (the sweep flags 5 of the 6 single-index forms).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/tensors/abstractblocktensor/abstractarray.jl	71.71% <100.00%> (+1.91%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

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