- DiffengineConeProgram: extract the shared x=0 matrix evaluation (q, d, A, b, P)
  into _extract_matrices, used by both from_problem and apply_parameters instead of
  duplicating it. Behavior unchanged.
- convert_symbolic_quad_form / _lower_symbolic_power: condense the verbose
  docstrings/comments; logic unchanged (all sparse/dense/parametric/Power branches
  are exercised).
- Drop the verbose DIFFENGINE explanatory comments in cone_matrix_stuffing.py and
  solving_chain.py.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Both ParamProb backends (ParamConeProg and DiffengineConeProgram) already return
concrete (q, d, A, b) from apply_parameters(), and the perspective aux problem is
parameter-free. Use that single call instead of branching on the program type and
hand-decoding ParamConeProg's embedded sparse tensors. Removes the isinstance
branch, the DiffengineConeProgram import, and the brittle reshape(order="F") decode.

A is kept sparse (sparse @ x_canon works and avoids embedding a dense constant in
the enclosing problem).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The diffengine cone backend's quad_form handling requires the dense-quadform
bindings shipped in the 0.5.x line.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The diffengine cone backend is tightly coupled to the sparsediffpy binding ABI
(dense make_quad_form), so pin the exact version rather than a range.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Remove the lb_tensor/ub_tensor placeholders (parametric bounds are not supported
on the diffengine path and nothing reads them), the unused param_id_to_size, and
the self-referential quad_obj attribute (the solver already keys quadratic-ness
off `problem.P is None`). No external readers of any of these.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Nothing reads it (apply_parameters reads p.value directly rather than going
through a param_id->col map). The other accessors stay: variables (read by the
HIGHS/Gurobi/Xpress interfaces), var_id_to_col (perspective_canon + split_solution),
and id_to_var (split_solution).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Drop the intra-method grouping blank lines in __init__, apply_parameters,
apply_restruct_mat, and from_problem. No behavior change.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Mirror the Jacobian pattern for the Hessian: capture the (fixed) lower-triangular
Hessian sparsity once in build() via init_hessian_coo_lower_tri +
get_problem_hessian_sparsity_coo, then on each re-solve fetch only the lower-tri
values (eval_hessian_vals_coo_lower_tri) and mirror off-diagonals to the full
symmetric P the cone solvers expect. Reuses existing bindings (no engine change).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The lazy-operand variant was confirmed unrelated to the matmul slowdown/segfault
(reverting left both unchanged: Murray ~15.5s cold compile, OptimalAdvertising
still segfaults). The slowdown/crash live in make_dense/sparse_left_matmul + the C
engine, invoked identically by both forms. Keep the simpler eager version;
convert_expr converts all args up front and convert_matmul consumes children.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>