t2s constant_mul drops a non-numeric scalar_wrapper factor: wrapper(x)*(trace(A)*det(A)) → 1*det(A)*tr(A)

[petlenz]

Summary

constant_mul::dispatch(tensor_to_scalar_mul) (src/numsim_cas/tensor_to_scalar/simplifier/tensor_to_scalar_simplifier_mul.cpp:31-44) silently drops a non-numeric scalar_wrapper LHS when multiplying into an existing product (reproduced at 0f6c0b9):

// x scalar symbol, A tensor
wrapper(x) * (trace(A) * det(A))   // → "1*det(A)*tr(A)" — x has vanished, coeff reset to 1

When lhs_val (numeric extraction) is nullopt, the code computes coeff{get_coefficient(mul,1)}, skips the coeff * *lhs_val merge, and never inserts m_lhs into the product. The promoted path x * (f*g) happens to route through mul_base::dispatch and is correct — which is why tests miss it; the bug fires whenever the wrapper is the visitor's LHS (e.g. (x*f)*g built in an order that puts the wrapper first, or internal simplifier calls).

Severity: high (silent factor loss). Confirmed. Related: #104 (mul×mul merge consolidation — this is a concrete drift instance of that risk).

Fix

In the !lhs_val branch, insert m_lhs as a child (push_or_combine(mul, m_lhs), merging with an existing scalar_wrapper child the way n_ary_add::dispatch(scalar_wrapper) does) instead of touching only the numeric coefficient.

Tests (lock-in)

TEST(T2sConstantMul, SymbolicWrapperFactorSurvives) {
  auto [x] = make_scalar_variable("x");
  auto [A] = make_tensor_variable(std::tuple{"A", 3, 2});
  auto f = trace(A) * det(A);
  auto e = /* wrapper-first product, e.g. via promote/explicit wrapper */ wrap(x) * f;
  // evaluate with x=2, A=diag(1,2,3): trace=6, det=6 → expect 72, not 36
  EXPECT_NE(to_string(e).find("x"), std::string::npos);
}

(Construct the wrapper-first path the same way the probe did — building tensor_to_scalar_scalar_wrapper(x) explicitly and multiplying it by an existing tensor_to_scalar_mul — so the regression can't dodge through the promoted route.)