[dance858]

Summary

• Bug: add evaluated both children's Hessians and then summed them. When the children share a node, evaluating the right child re-evaluates that node with different weights. That overwrites Hessian values the left child still needed before the sum reads them. add(log x, exp(log x)) gave -2 where -4 is correct.
• Fix: the Hessian sum is now filled incrementally with sum_n_matrices_fill_zero / sum_n_matrices_fill_add, so each child's Hessian is added right after it is evaluated. The pattern comes from sum_n_matrices_alloc with n = 2, and add still needs no struct of its own.
• Jacobians: the Jacobian path is unchanged. A shared node's Jacobian doesn't depend on weights, so re-evaluating it writes the same values.
• Pin: sparsediff_linalg moves to cba21ac (library Adds indexing atom implementation and other bindings #20, which adds fill_zero / fill_add).

Other atoms that evaluate several children and sum afterwards (multiply/matmul via sum_4, problem.c's Lagrange Hessian) likely have the same issue with shared nodes. They aren't changed here.

Tests

• New test_wsum_hess_add_shared_subexpr checks add(log x, exp(log x)) and add(exp(log x), log x) against the numerical Hessian. Both use a single shared log x node.
• The test fails on the old add.c (analytical -2, numerical -4) and passes with the fix.
• All 301 tests pass in Debug and Release.

🤖 Generated with Claude Code