Refresh composite param_source subtrees on parameter updates

[Transurgeon]

Problem

A param_source subtree is evaluated by the owning atom's gated recursion, not the main forward walk, so expr_set_needs_refresh never reaches nodes inside a composite source. Gated nodes there — promote, nested mults, cached matmul coefficients — keep serving values cached at construction after problem_update_params.

The failing shapes are common, not exotic (each verified with a standalone repro):

lowered shape	example cvxpy origin	result before this PR
vector_mult(promote(p), A) feeding left_matmul	(p * A) @ x	stale A on every re-solve
vector_mult(promote(g), Sig) feeding quad_form	quad_form(x, g * Sig)	stale Q
division/scaling chains feeding quad_form	quad_over_lin(x, p) (canonicalized to I/p)	stale Q
two nested gated levels	((2*p) * A) @ x	stale A

Shapes that were already correct and stay untouched: bare-Parameter sources (refreshed by problem_update_params' memcpy), purely additive composites (p1+p2 — ungated nodes recurse unconditionally), and computed multipliers of variables ((2*p)*x — the gated node sits in the main tree, which the walker reaches).

Fix

One line per source-owning atom (scalar_mult, vector_mult, left_matmul — which also serves right_matmul — convolve, kron, quad_form): mark the side subtree dirty at the moment the atom is about to re-evaluate it,

if (node->param_source != NULL && node->needs_parameter_refresh)
{
    expr_set_needs_refresh(node->param_source);   /* new */
    node->param_source->forward(node->param_source, NULL);
}

The owner's flag is already set by the existing walker (it lives in the main tree), so the mark happens exactly once per parameter update; nested gated levels recurse for free because each level applies the same line to its own source. No struct changes, no walker changes, no behavior change for one-shot solves or bare-parameter sources.

Tests

tests/problem/test_param_source_refresh.h pins the three failing shapes (composite matmul coefficient, composite quad matrix, two nested gated levels), asserting constraint values and Jacobian/gradient values across an update. Each fails without the fix (e.g. stale 10.0 where 50.0 is expected) and passes with it. Full suite: 417/417.

Downstream validation (cvxpy pr-a/b/c stack, cached re-solve paths): the cvxpy-side canaries (quad_over_lin symbolic quad matrix, scaled parameter coefficient) and all four repro shapes above go from stale to fresh on this branch with no other changes; full cvxpy test suite green (modulo pre-existing, unrelated failures verified identical on other builds).

Relation to #104

This supersedes the first commit of #104 (param-source refresh propagation) with a smaller, atom-local change: no param_source pointer hoisted into the base expr struct. #104's second commit (retaining registered parameter nodes) is deliberately not included — it is an independent API-soundness hardening that can be its own PR if wanted.

🤖 Generated with Claude Code

[dance858]

Beautiful solution. I could stare at it for hours.