einsum: delegate nesting-preserving products to the expression layer; drop per-tile reference kernels

[evaleev]

Stacked on #562 (evaleev/feature/general-product-expr). Narrows einsum so that, by default, only type-mutating products (the ToT·ToT→T denest reduction) remain its own; every nesting-preserving product now flows through the general-product expression engine.

Commits

1. einsum: delegate nesting-preserving products to the expression layer

• Fused broadcast C(h,e) = A(h) · B(h,e) (one operand entirely fused, no contraction): drops the replicate_array() expand-then-Hadamard and delegates directly. The engine evaluates it natively once the fully-fused operand leads — its folded left tile is rank-0, restored by a synthetic unit left-external mode. ContEngine::synthetic_unit_left_external() now fires for a rank-0 left operand (broadcast), not only a rank-0 result (no-external). einsum orders the fully-fused operand first and applies any inner result permutation afterward.
• No-external C(h;…) = A(h,i;…) · B(h,i;…) (fused + contracted, no free index): now defaults to the expression layer (the existing no-external synthetic-unit mode). The legacy local kernel is retained only behind einsum_legacy_subworld() as the cross-check oracle, mirroring the generalized-subworld path.
• Tests: dense fused broadcast (both operand orders); ToT no-external with inner Hadamard and inner contraction.

2. tests: guard assign_subblock REQUIRE_THROW on TA_ASSERT_POLICY

• That throw is enforced by a TA_ASSERT, so it only happens under TA_ASSERT_THROW. Under the Release default (TA_ASSERT_IGNORE) the assert is a no-op and the REQUIRE_THROW failed; guard it to only expect the throw when assertions actually throw.

3. tests: remove the per-tile general-product reference kernels

• contraction_helpers.h (kernels::{s_t_t,t_s_t,t_t_t,t_tot_tot,tot_t_tot,tot_tot_tot}_contract_) and its dedicated tests were scaffolding used only to develop/cross-check einsum; the expression-layer engine supersedes them and no production code references them. Removes the header, the generated test files, their CMake entries, the stale einsum.cpp include, and the generate_answers.ipynb generator.

Validation

einsum (einsum_manual, einsum_tot, einsum_tot_t, einsum_tiledarray, einsum_eigen) + general_product_suite — 102 cases green post-nuke.

[Copilot]

Pull request overview

This PR narrows einsum() so that nesting-preserving products (including fused broadcast and “no-external” fused+contracted reductions) are evaluated by the expression-layer general-product engine by default, and removes the legacy per-tile reference kernels/tests that were previously used as a development oracle.

Changes:

• Route fused-broadcast and no-external (nesting-preserving) products through the expression engine (with operand ordering + canonical-eval + final permutation where needed).
• Extend ContEngine’s synthetic-unit handling to cover rank-0 left operand folds (broadcast) in addition to rank-0 result folds (no-external).
• Remove contraction_helpers.h and its associated generated tests/scaffolding (plus related build/test wiring).

Reviewed changes

Copilot reviewed 11 out of 15 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/t_s_t_contract_.cpp	Removed generated tests for legacy per-tile contraction helper kernels.
tests/s_t_t_contract_.cpp	Removed generated tests for legacy per-tile contraction helper kernels.
tests/general_product.cpp	Added coverage for dense fused-broadcast and ToT no-external cases; removed debug try/catch noise.
tests/expressions_impl.h	Guarded a throw-expectation that only applies when TA_ASSERT is configured to throw.
tests/einsum.cpp	Dropped now-removed legacy header include; formatting-only adjustments in touched hunks.
tests/contraction_helpers.cpp	Removed tests for legacy helper utilities tied to the deleted header.
tests/CMakeLists.txt	Removed deleted legacy test sources from the test build.
src/TiledArray/expressions/contraction_helpers.h	Removed legacy helper header and its old einsum implementation.
src/TiledArray/expressions/cont_engine.h	General-product engine: synthetic-unit mode now also supports rank-0 left operand folds (broadcast).
src/TiledArray/einsum/tiledarray.h	einsum() delegates fused-broadcast + no-external nesting-preserving products to the expression layer (with canonicalization and post-permute).
src/CMakeLists.txt	Removed deleted header from the installed/source header list.
bin/generate_answers.ipynb	Removed obsolete generator notebook for the deleted reference-kernel tests.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.