diff --git a/levels/thi/areas/2/collision.inc.c b/levels/thi/areas/2/collision.inc.c
index dce7d94a0..71f500740 100644
--- a/levels/thi/areas/2/collision.inc.c
+++ b/levels/thi/areas/2/collision.inc.c
@@ -355,10 +355,10 @@ const Collision thi_seg7_area_2_collision[] = {
     COL_VERTEX(-2457, -3788, -2457),
     COL_VERTEX(2458, -3788, 2458),
     COL_VERTEX(-2457, -3788, 2458),
-    COL_VERTEX(2335, -3788, 61),
-    COL_VERTEX(2335, -3788, -859),
-    COL_VERTEX(1690, -3788, -859),
-    COL_VERTEX(1690, -3788, 61),
+    COL_VERTEX(2335, -3789, 61),
+    COL_VERTEX(2335, -3789, -859),
+    COL_VERTEX(1690, -3789, -859),
+    COL_VERTEX(1690, -3789, 61),
     COL_VERTEX(1997, -921, 2458),
     COL_VERTEX(-1535, -921, -2457),
     COL_VERTEX(-2457, -921, -2457),
diff --git a/src/boot/main.c b/src/boot/main.c
index c4d5b9e6c..b22639b76 100644
--- a/src/boot/main.c
+++ b/src/boot/main.c
@@ -333,6 +333,33 @@ void check_cache_emulation() {
     __osRestoreInt(saved);
 }
 
+/**
+ * Increment the first and last values of the stack.
+ * If they're different, that means an error has occured, so trigger a crash.
+*/
+#ifdef DEBUG
+void check_stack_validity(void) {
+    gIdleThreadStack[0]++;
+    gIdleThreadStack[THREAD1_STACK - 1]++;
+    assert(gIdleThreadStack[0] == gIdleThreadStack[THREAD1_STACK - 1], "Thread 1 stack overflow.")
+    gThread3Stack[0]++;
+    gThread3Stack[THREAD3_STACK - 1]++;
+    assert(gThread3Stack[0] == gThread3Stack[THREAD3_STACK - 1], "Thread 3 stack overflow.")
+    gThread4Stack[0]++;
+    gThread4Stack[THREAD4_STACK - 1]++;
+    assert(gThread4Stack[0] == gThread4Stack[THREAD4_STACK - 1], "Thread 4 stack overflow.")
+    gThread5Stack[0]++;
+    gThread5Stack[THREAD5_STACK - 1]++;
+    assert(gThread5Stack[0] == gThread5Stack[THREAD5_STACK - 1], "Thread 5 stack overflow.")
+#if ENABLE_RUMBLE
+    gThread6Stack[0]++;
+    gThread6Stack[THREAD6_STACK - 1]++;
+    assert(gThread6Stack[0] == gThread6Stack[THREAD6_STACK - 1], "Thread 6 stack overflow.")
+#endif
+}
+#endif
+
+
 extern void crash_screen_init(void);
 extern OSViMode VI;
 void thread3_main(UNUSED void *arg) {
@@ -376,16 +403,33 @@ void thread3_main(UNUSED void *arg) {
         gIsConsole = TRUE;
         gBorderHeight = BORDER_HEIGHT_CONSOLE;
     }
+#ifdef DEBUG
+    gIdleThreadStack[0] = 0;
+    gIdleThreadStack[THREAD1_STACK - 1] = 0;
+    gThread3Stack[0] = 0;
+    gThread3Stack[THREAD3_STACK - 1] = 0;
+    gThread4Stack[0] = 0;
+    gThread4Stack[THREAD4_STACK - 1] = 0;
+    gThread5Stack[0] = 0;
+    gThread5Stack[THREAD5_STACK - 1] = 0;
+#if ENABLE_RUMBLE
+    gThread6Stack[0] = 0;
+    gThread6Stack[THREAD6_STACK - 1] = 0;
+#endif
+#endif
 
-    create_thread(&gSoundThread, THREAD_4_SOUND, thread4_sound, NULL, gThread4Stack + 0x2000, 20);
+    create_thread(&gSoundThread, THREAD_4_SOUND, thread4_sound, NULL, gThread4Stack + THREAD4_STACK, 20);
     osStartThread(&gSoundThread);
 
-    create_thread(&gGameLoopThread, THREAD_5_GAME_LOOP, thread5_game_loop, NULL, gThread5Stack + 0x2000, 10);
+    create_thread(&gGameLoopThread, THREAD_5_GAME_LOOP, thread5_game_loop, NULL, gThread5Stack + THREAD5_STACK, 10);
     osStartThread(&gGameLoopThread);
 
     while (TRUE) {
         OSMesg msg;
         osRecvMesg(&gIntrMesgQueue, &msg, OS_MESG_BLOCK);
+#ifdef DEBUG
+        check_stack_validity();
+#endif
         switch ((uintptr_t) msg) {
             case MESG_VI_VBLANK:
                 handle_vblank();
@@ -526,7 +570,7 @@ void thread1_idle(UNUSED void *arg) {
     osViSetSpecialFeatures(OS_VI_DITHER_FILTER_ON);
     osViSetSpecialFeatures(OS_VI_GAMMA_OFF);
     osCreatePiManager(OS_PRIORITY_PIMGR, &gPIMesgQueue, gPIMesgBuf, ARRAY_COUNT(gPIMesgBuf));
-    create_thread(&gMainThread, THREAD_3_MAIN, thread3_main, NULL, gThread3Stack + 0x2000, 100);
+    create_thread(&gMainThread, THREAD_3_MAIN, thread3_main, NULL, gThread3Stack + THREAD3_STACK, 100);
     osStartThread(&gMainThread);
 
     osSetThreadPri(NULL, 0);
@@ -564,6 +608,6 @@ void main_func(void) {
 #ifdef ISVPRINT
     osInitialize_fakeisv();
 #endif
-    create_thread(&gIdleThread, THREAD_1_IDLE, thread1_idle, NULL, gIdleThreadStack + 0x800, 100);
+    create_thread(&gIdleThread, THREAD_1_IDLE, thread1_idle, NULL, gIdleThreadStack + THREAD1_STACK, 100);
     osStartThread(&gIdleThread);
 }
diff --git a/src/buffers/buffers.c b/src/buffers/buffers.c
index 5d382702f..5a8eefa2d 100644
--- a/src/buffers/buffers.c
+++ b/src/buffers/buffers.c
@@ -12,12 +12,12 @@
 ALIGNED8 u8 gDecompressionHeap[0xD000];
 ALIGNED16 u8 gAudioHeap[DOUBLE_SIZE_ON_64_BIT(AUDIO_HEAP_SIZE)];
 
-ALIGNED8 u8 gIdleThreadStack[0x800];
-ALIGNED8 u8 gThread3Stack[0x2000];
-ALIGNED8 u8 gThread4Stack[0x2000];
-ALIGNED8 u8 gThread5Stack[0x2000];
+ALIGNED8 u8 gIdleThreadStack[THREAD1_STACK];
+ALIGNED8 u8 gThread3Stack[THREAD3_STACK];
+ALIGNED8 u8 gThread4Stack[THREAD4_STACK];
+ALIGNED8 u8 gThread5Stack[THREAD5_STACK];
 #if ENABLE_RUMBLE
-ALIGNED8 u8 gThread6Stack[0x2000];
+ALIGNED8 u8 gThread6Stack[THREAD6_STACK];
 #endif
 // 0x400 bytes
 #if UNF
diff --git a/src/buffers/buffers.h b/src/buffers/buffers.h
index f063a023c..82b54b4eb 100644
--- a/src/buffers/buffers.h
+++ b/src/buffers/buffers.h
@@ -8,16 +8,23 @@
 #include "game/game_init.h"
 #include "config.h"
 
+#define THREAD1_STACK 0x100
+#define THREAD2_STACK 0x800
+#define THREAD3_STACK 0x200
+#define THREAD4_STACK 0x2000
+#define THREAD5_STACK 0x2000
+#define THREAD6_STACK 0x400
+
 extern u8 gDecompressionHeap[];
 
 extern u8 gAudioHeap[];
 
-extern u8 gIdleThreadStack[];
-extern u8 gThread3Stack[];
-extern u8 gThread4Stack[];
-extern u8 gThread5Stack[];
+extern u8 gIdleThreadStack[THREAD1_STACK];
+extern u8 gThread3Stack[THREAD3_STACK];
+extern u8 gThread4Stack[THREAD4_STACK];
+extern u8 gThread5Stack[THREAD5_STACK];
 #if ENABLE_RUMBLE
-extern u8 gThread6Stack[];
+extern u8 gThread6Stack[THREAD6_STACK];
 #endif
 
 extern u8 gGfxSPTaskYieldBuffer[];
diff --git a/src/buffers/gfx_output_buffer.c b/src/buffers/gfx_output_buffer.c
index 770c4725e..08675bb10 100644
--- a/src/buffers/gfx_output_buffer.c
+++ b/src/buffers/gfx_output_buffer.c
@@ -1,4 +1,4 @@
 #include <ultra64.h>
 #include "gfx_output_buffer.h"
 
-u64 gGfxSPTaskOutputBuffer[0x2fc0];
+u64 gGfxSPTaskOutputBuffer[0x2000];
diff --git a/src/buffers/gfx_output_buffer.h b/src/buffers/gfx_output_buffer.h
index 8f6ded2f9..abf08afb6 100644
--- a/src/buffers/gfx_output_buffer.h
+++ b/src/buffers/gfx_output_buffer.h
@@ -3,6 +3,6 @@
 
 #include <PR/ultratypes.h>
 
-extern u64 gGfxSPTaskOutputBuffer[0x2fc0];
+extern u64 gGfxSPTaskOutputBuffer[0x2000];
 
 #endif // GFX_OUTPUT_BUFFER_H
diff --git a/src/engine/surface_load.c b/src/engine/surface_load.c
index aca97f52a..e4a505b9e 100644
--- a/src/engine/surface_load.c
+++ b/src/engine/surface_load.c
@@ -119,7 +119,6 @@ static void clear_static_surfaces(void) {
  */
 static void add_surface_to_cell(s32 dynamic, s32 cellX, s32 cellZ, struct Surface *surface) {
     struct SurfaceNode *list;
-    s32 priority;
     s32 sortDir = 1; // highest to lowest, then insertion order (water and floors)
     s32 listIndex;
 
@@ -135,7 +134,6 @@ static void add_surface_to_cell(s32 dynamic, s32 cellX, s32 cellZ, struct Surfac
         sortDir = 0; // insertion order
     }
 
-    s32 surfacePriority = surface->upperY * sortDir;
 
     struct SurfaceNode *newNode = alloc_surface_node(dynamic);
     newNode->surface = surface;
@@ -157,14 +155,18 @@ static void add_surface_to_cell(s32 dynamic, s32 cellX, s32 cellZ, struct Surfac
     }
 
     // Loop until we find the appropriate place for the surface in the list.
-    while (list->next != NULL) {
-        priority = list->next->surface->upperY * sortDir;
+    if (listIndex == SPATIAL_PARTITION_WATER) {
+        s32 surfacePriority = surface->upperY * sortDir;
+        s32 priority;
+        while (list->next != NULL) {
+            priority = list->next->surface->upperY * sortDir;
 
-        if (surfacePriority > priority) {
-            break;
+            if (surfacePriority > priority) {
+                break;
+            }
+
+            list = list->next;
         }
-
-        list = list->next;
     }
 
     newNode->next = list->next;
@@ -269,7 +271,15 @@ static struct Surface *read_surface_data(TerrainData *vertexData, TerrainData **
 
     find_vector_perpendicular_to_plane(n, v[0], v[1], v[2]);
 
-    vec3f_normalize(n);
+    f32 mag = (sqr(n[0]) + sqr(n[1]) + sqr(n[2]));
+    // This will never need to be run for custom levels because Fast64 does this step before exporting.
+#ifdef ENABLE_VANILLA_LEVEL_SPECIFIC_CHECKS
+    if (mag < NEAR_ZERO) {
+        return NULL;
+    }
+#endif
+    mag = 1.0f / sqrtf(mag);
+    vec3_mul_val(n, mag);
 
     struct Surface *surface = alloc_surface(dynamic);
 
diff --git a/src/game/crash_screen.c b/src/game/crash_screen.c
index 0913e6bae..3690b32f8 100644
--- a/src/game/crash_screen.c
+++ b/src/game/crash_screen.c
@@ -11,6 +11,7 @@
 #include "main.h"
 #include "debug.h"
 #include "rumble_init.h"
+#include "buffers/buffers.h"
 
 #include "sm64.h"
 
@@ -80,7 +81,7 @@ extern far char *find_function_in_stack(u32 *sp);
 
 struct {
     OSThread thread;
-    u64 stack[0x800 / sizeof(u64)];
+    u64 stack[THREAD2_STACK / sizeof(u64)];
     OSMesgQueue mesgQueue;
     OSMesg mesg;
     u16 *framebuffer;
diff --git a/src/game/rumble_init.c b/src/game/rumble_init.c
index b164798e0..2ebf75d0e 100644
--- a/src/game/rumble_init.c
+++ b/src/game/rumble_init.c
@@ -261,7 +261,7 @@ void cancel_rumble(void) {
 
 void create_thread_6(void) {
     osCreateMesgQueue(&gRumbleThreadVIMesgQueue, gRumbleThreadVIMesgBuf, 1);
-    osCreateThread(&gRumblePakThread, THREAD_6_RUMBLE, thread6_rumble_loop, NULL, gThread6Stack + 0x2000, 30);
+    osCreateThread(&gRumblePakThread, THREAD_6_RUMBLE, thread6_rumble_loop, NULL, gThread6Stack + 0x400, 30);
     osStartThread(&gRumblePakThread);
 }