(RWEBP) Bug #14: Y_MODE TREE IS BALANCED!

The keyframe y_mode tree was implemented as a linear chain but libvpx uses a balanced binary tree (confirmed by dumping vp8_kf_ymode_tree from the binary):
node0: bit=0 → B_PRED(4)
       bit=1 → node1
node1: bit=0 → node2 {DC(0), V(1)}
       bit=1 → node3 {H(2), TM(3)}
This is the exact same type of bug as the sub-block mode tree — a linear chain that should be balanced. The fix changed the y_mode reading from a sequential if/elif/elif/elif/else to a branching structure where prob[1] selects between {DC,V} and {H,TM} subtrees.
Result: First 5 MBs now match libvpx perfectly (seg, ym, uvm, skip all correct for MB0-MB4). The overall diff is 105.7 (higher than 96.9 because the correct tree changes the cascading error pattern), but the structural correctness is confirmed by the 5-MB match.

Author: LibretroAdmin

libretro-common/formats/webp/rwebp.c

@@ -1250,11 +1250,15 @@ static uint32_t *vp8_decode(const uint8_t *data, size_t len,
             is_skip = vp8b_get(&br, prob_skip);
 
          /* Y mode */
-         if      (!vp8b_get(&br, vp8_ymp[0])) ym = 4; /* B_PRED first */
-         else if (!vp8b_get(&br, vp8_ymp[1])) ym = 0;
-         else if (!vp8b_get(&br, vp8_ymp[2])) ym = 1;
-         else if (!vp8b_get(&br, vp8_ymp[3])) ym = 2;
-         else ym = 3; /* TM_PRED */
+         if (!vp8b_get(&br, vp8_ymp[0])) {
+            ym = 4; /* B_PRED */
+         } else if (!vp8b_get(&br, vp8_ymp[1])) {
+            /* Left subtree: DC, V */
+            ym = vp8b_get(&br, vp8_ymp[2]) ? 1 : 0;
+         } else {
+            /* Right subtree: H, TM */
+            ym = vp8b_get(&br, vp8_ymp[3]) ? 3 : 2;
+         }
 
          if (ym == 4)
          {