Model the DMG window LCD desync (mealybug wewx 521->83)

When a window's WX==position+7 activation dot is masked by a mid-line
LCDC.5-off pulse, the window still activates as the pulse re-enables it,
rendered as if it had triggered at the match dot. This is SameBoy's
WX==position+6 + lcd_x-- horizontal desync, generalised: our LCDC-pulse
commit lands a fixed number of dots off the pixel position axis, so a
literal +6 comparison misfires. Instead we record the +7 match position
reached while the window is disabled and, on re-enable, activate and
rewind the output back to it when it is within DESYNC_MAX_GAP dots. That
reproduces hardware's "activate iff the match is at or just before the
re-enable" across every affected line, so m3_lcdc_win_en_change_multiple_wx
drops from 521 to 83 diff pixels with no change to the other window tests.

Adds WX-just-written tracking to GpuRegisterValues (the desync is
suppressed for one machine cycle after a WX write, SameBoy wx_just_changed).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J1wLWscyUGS7CFwc3CjJR8

Author: trekawek

core/src/main/java/eu/rekawek/coffeegb/core/gpu/GpuRegisterValues.java

@@ -27,6 +27,13 @@ public class GpuRegisterValues implements AddressSpace, Serializable, Originator
 
     private final int[] pendingMixValues = new int[GpuRegister.values().length];
 
+    // ticks remaining in which WX counts as "just written": the DMG's WX==position+6
+    // window desync is suppressed for one machine cycle after a WX write (SameBoy
+    // wx_just_changed)
+    private int wxJustChangedTicks;
+
+    private static final int WX_CHANGE_TICKS = 2;
+
     private boolean gbc;
 
     public void setGbc(boolean gbc) {
@@ -59,6 +66,13 @@ void tickConflicts() {
             mixValues[reg.ordinal()] = pendingMixValues[reg.ordinal()];
             pendingMixValues[reg.ordinal()] = -1;
         }
+        if (wxJustChangedTicks > 0) {
+            wxJustChangedTicks--;
+        }
+    }
+
+    public boolean isWxJustChanged() {
+        return wxJustChangedTicks > 0;
     }
 
     private static final GpuRegister[] PALETTE_REGISTERS =
@@ -85,6 +99,9 @@ public void setByte(int address, int value) {
             if (!gbc && (reg == GpuRegister.BGP || reg == GpuRegister.OBP0 || reg == GpuRegister.OBP1)) {
                 pendingMixValues[reg.ordinal()] = values[reg.ordinal()] | value;
             }
+            if (reg == GpuRegister.WX) {
+                wxJustChangedTicks = WX_CHANGE_TICKS;
+            }
             values[reg.ordinal()] = value;
         }
     }
@@ -110,7 +127,7 @@ private static GpuRegister fromAddress(int address) {
 
     @Override
     public Memento<GpuRegisterValues> saveToMemento() {
-        return new GpuRegisterValuesMemento(values.clone(), mixValues.clone(), pendingMixValues.clone());
+        return new GpuRegisterValuesMemento(values.clone(), mixValues.clone(), pendingMixValues.clone(), wxJustChangedTicks);
     }
 
     @Override
@@ -129,9 +146,11 @@ public void restoreFromMemento(Memento<GpuRegisterValues> memento) {
             java.util.Arrays.fill(this.mixValues, -1);
             java.util.Arrays.fill(this.pendingMixValues, -1);
         }
+        this.wxJustChangedTicks = mem.wxJustChangedTicks;
     }
 
-    private record GpuRegisterValuesMemento(int[] values, int[] mixValues, int[] pendingMixValues)
+    private record GpuRegisterValuesMemento(int[] values, int[] mixValues, int[] pendingMixValues,
+                                            int wxJustChangedTicks)
             implements Memento<GpuRegisterValues> {
     }
 }

core/src/main/java/eu/rekawek/coffeegb/core/gpu/phase/PixelTransfer.java

@@ -74,6 +74,10 @@ public class PixelTransfer implements GpuPhase, Serializable, Originator<PixelTr
     // rolled back at commit so the activation behaves as if it happened at the match
     private int windowPendingPos;
 
+    // position of a WX==position+7 match that landed while the window was disabled by a
+    // mid-line pulse, awaiting the window re-enable to activate (DMG desync); -1 = none
+    private int windowCatchUpPos = -1;
+
     // whether the window already activated on this line: a RE-activation (after a
     // mid-line window drop, e.g. an LCDC.5 toggle) restarts its fetch one T-cycle
     // later than the line's first activation (m3_lcdc_win_en_change_multiple)
@@ -158,6 +162,7 @@ public PixelTransfer start(int scyLatch, int extraEntryDelay) {
         machineActive = true;
         position = -16;
         windowActivatedThisLine = false;
+        windowCatchUpPos = -1;
         machineStall = 0;
         fifo.startLine();
         window = false;
@@ -305,6 +310,19 @@ public boolean tick() {
             windowBeingFetched = false;
         }
 
+        // record a WX==position+7 match that lands while the window is disabled by a
+        // mid-line pulse: the window will "catch up" when it is re-enabled (see the
+        // desync branch below). Only for regular WX>=7 positions so the discard-phase
+        // WX=0..6 activation is untouched.
+        if (!gbc && !window && !lcdc.isWindowDisplay() && !windowActivatedThisLine
+                && r.get(LY) >= r.get(WY)) {
+            int wxNow = r.get(WX);
+            if (wxNow >= 7 && wxNow < 166 && !r.isWxJustChanged()
+                    && wxNow == ((position + 7) & 0xff)) {
+                windowCatchUpPos = position;
+            }
+        }
+
         // window activation check (SameBoy model): the comparison wraps in 8 bits, so
         // WX values 0..6 can trigger during the discard phase; WX=0 has its own window,
         // and the CGB also accepts WX=166. The window line counter increments at the
@@ -323,7 +341,28 @@ public boolean tick() {
             } else {
                 activate = false;
             }
-            if (activate && windowPendingTicks == 0) {
+            if (!gbc && windowCatchUpPos >= 0 && windowPendingTicks == 0
+                    && !windowActivatedThisLine
+                    && position - windowCatchUpPos <= DESYNC_MAX_GAP) {
+                // DMG LCD-PPU horizontal desync: the WX==position+7 match dot passed while
+                // the window was disabled (a mid-line LCDC.5-off pulse); the window catches
+                // up on re-enable, rendered as if it had triggered at the match dot. This is
+                // SameBoy's WX==position+6 + lcd_x-- generalised to our pulse/position phase
+                // (mealybug m3_lcdc_win_en_change_multiple_wx). DESYNC_MAX_GAP encodes the
+                // phase between our LCDC-pulse commit and the pixel position axis.
+                window = true;
+                windowBeingFetched = true;
+                windowLineCounter++;
+                for (int p = windowCatchUpPos; p < position; p++) {
+                    fifo.rewindOnePixel();
+                }
+                position = windowCatchUpPos;
+                windowCatchUpPos = -1;
+                fifo.clear();
+                fetcher.startWindow();
+                fetcher.advance(position, true, windowLineCounter, false);
+                windowActivatedThisLine = true;
+            } else if (activate && windowPendingTicks == 0) {
                 windowPendingTicks = 1;
                 windowPendingWx = wx;
                 windowPendingPos = position;
@@ -384,6 +423,11 @@ public boolean tick() {
     }
 
 
+    // gap (in dots) between our LCDC-pulse re-enable commit and the pixel position axis:
+    // a window whose WX match was masked by a pulse catches up on re-enable only if the
+    // match is within this many dots of the re-enable position (mealybug wewx)
+    private static final int DESYNC_MAX_GAP = 3;
+
     private void objTick() {
         if (objStep <= 1) {
             fetcher.advance(position, window, windowLineCounter, true);

core/src/test/java/eu/rekawek/coffeegb/core/integration/mealybug/MealybugRomTest.java

@@ -42,7 +42,7 @@ public class MealybugRomTest {
             Map.entry("m3_lcdc_tile_sel_change.gb", 0),
             Map.entry("m3_lcdc_tile_sel_win_change.gb", 192),
             Map.entry("m3_lcdc_win_en_change_multiple.gb", 0),
-            Map.entry("m3_lcdc_win_en_change_multiple_wx.gb", 521),
+            Map.entry("m3_lcdc_win_en_change_multiple_wx.gb", 83),
             Map.entry("m3_lcdc_win_map_change.gb", 92),
             Map.entry("m3_obp0_change.gb", 0),
             Map.entry("m3_scx_high_5_bits.gb", 0),