Refine line search convergence criteria.

The line search now requires a negative cost (improvement) in addition to a small derivative to declare convergence, preventing premature termination when no actual improvement has been made.

Follows the proposal in github.com/google-deepmind/mujoco_warp/pull/1471

PiperOrigin-RevId: 941579503
Change-Id: I8fcd20f7b959e50d77cd5d6de0a3c6d95f86b9e1

Author: yuvaltassa

src/engine/engine_solver.c

@@ -1855,7 +1855,7 @@ static mjtNum PrimalSearch(mjPrimalContext* ctx, mjtNum tolerance, mjtNum ls_ite
   PrimalEval(ctx, &p1);
 
   // check for initial convergence
-  if (mju_abs(p1.deriv[0]) < gtol) {
+  if (mju_abs(p1.deriv[0]) < gtol && (p1.alpha == 0 || p1.cost < 0)) {
     if (p1.alpha == 0) {
       ctx->LSresult = 2;  // no improvement, initial convergence
     } else {
@@ -1916,7 +1916,7 @@ static mjtNum PrimalSearch(mjPrimalContext* ctx, mjtNum tolerance, mjtNum ls_ite
     PrimalEval(ctx, &p1);
 
     // check for convergence
-    if (mju_abs(p1.deriv[0]) < gtol) {
+    if (mju_abs(p1.deriv[0]) < gtol && p1.cost < 0) {
       ctx->LSslope = mju_abs(p1.deriv[0])*slopescl;
       *improvement = -p1.cost;
       return p1.alpha;                          // SUCCESS