tensor_inv wrapper ctor lacks rank gate (rank-3 silently constructible)

[petlenz]

Context

Surfaced in code review of PR #291 (#287 rank-4 inv-diff scalar-arg case).

The inv() factory in include/numsim_cas/tensor/tensor_functions.h:467 rejects rank ≠ 2 and rank ≠ 4 at construction with a clear invalid_expression_error. The tensor_inv wrapper constructor in include/numsim_cas/tensor/wrappers/tensor_inv.h:14-85, however, runs no rank check itself — it only branches on r == 2 / r == 4 for space propagation and silently leaves the space empty otherwise.

So make_expression<tensor_inv>(rank3_tensor) (bypassing the factory) builds a "valid-looking" rank-3 tensor_inv AST node that misbehaves anywhere downstream:

• Visitors (differentiation, evaluator, printer) hit the wrong rank path.
• The space-propagation block silently produces a rank-3 tensor_inv with no space tag and the wrong algebraic implications.
• The diff visitor in Rank-4 tensor_inv in scalar-arg diff visitor (#287) #291 has a belt-and-braces throw that catches it for that particular visitor — but every other consumer is silently broken.

Fix

Mirror the factory's rank gate in the wrapper ctor:

template <typename Expr>
explicit tensor_inv(Expr &&_expr)
    : base(std::forward<Expr>(_expr), _expr.get().dim(), _expr.get().rank()) {
  if (this->rank() != 2 && this->rank() != 4) {
    throw invalid_expression_error(
        "tensor_inv: only rank-2 and rank-4 tensors are supported (got rank " +
        std::to_string(this->rank()) + ")");
  }
  // ... existing space and algebra-assumption propagation
}

Throws are the codebase convention here (matches inv() factory's invalid_expression_error style) and — unlike assert — fires in Release builds too.

Test plan

• Lock-in test: EXPECT_THROW(make_expression<tensor_inv>(rank3_T), invalid_expression_error).
• The belt-and-braces throw in tensor_differentiation_wrt_scalar.cpp:248-253 becomes unreachable; either:
  (a) Keep it with a comment "should be caught by wrapper ctor; double-guard for direct base-class construction".
  (b) Replace with assert(r == 2 || r == 4).
• The corresponding test TensorDiffWrtScalarTest.InvRank3ThrowsNotImplemented becomes redundant; can either be kept as a defense-in-depth witness or replaced by a wrapper-level direct test.

Related

• Factory gate this would mirror: tensor_functions.h:467 (inv() accepts non-rank-2 inputs even though comment says "Rank-2 only"; rank-4 path constructs a symbolic node that won't evaluate #192, Should tensor_inv support rank-4 inverses for the algorithmic-tangent use case? #248).
• Surfaced by: code review of Rank-4 tensor_inv in scalar-arg diff visitor (#287) #291.
• Belt-and-braces throw whose primary purpose disappears with this fix: src/numsim_cas/tensor/visitors/tensor_differentiation_wrt_scalar.cpp:248-253.