Remove temporary V38 build workflow

src/main.zig

@@ -1048,11 +1048,17 @@ comptime {
 pub export fn DllMain(
     _: ?*anyopaque,
     reason: u32,
-    _: ?*anyopaque,
+    reserved: ?*anyopaque,
 ) callconv(WINAPI) std.os.windows.BOOL {
     switch (reason) {
         1 => install(),
-        0 => uninstall(),
+        0 => {
+            // ExitFix: when lpReserved is non-null, Windows is terminating the
+            // whole process. Do not call into D3D9/COM or other DLLs from
+            // DLL_PROCESS_DETACH during global teardown; the OS will reclaim
+            // those resources. Keep full uninstall for explicit FreeLibrary.
+            if (reserved == null) uninstall();
+        },
         else => {},
     }
     return @enumFromInt(1);

src/outline/d3d9_hook.zig

@@ -57,6 +57,14 @@ var orig_reset: usize = 0;
 var d3d9_vtable: ?[*]usize = null;
 var hooks_installed: bool = false;
 
+/// ExitFix: once shutdown begins, no code path may reinstall D3D9 hooks.
+var shutdown_started: bool = false;
+
+pub fn beginShutdown() void {
+    shutdown_started = true;
+}
+
+
 // Sticky diagnostics for the in-game OutlineDebug() command.
 pub var debug_endscene_seen: bool = false;
 pub var debug_dip_seen: bool = false;
@@ -114,6 +122,7 @@ pub var debug_late_rehook_succeeded: bool = false;
 /// If the current entries are no longer ours, chain whatever is there now as
 /// the new originals, then patch the three entries again.
 pub fn lateRehookIfLost() bool {
+    if (shutdown_started) return false;
     debug_late_rehook_attempted = true;
 
     const cur = getD3D9VTable() orelse return false;
@@ -732,34 +741,45 @@ const jfa_prop_src =
     "mov oC0.xy, r8.xy\n" ++
     "mov oC0.zw, c1.xx\n";
 
-/// V32 POLISH: hard 3px outline, no feather.
-/// c0 = (screen_width, screen_height, radius_px=3, 0).
-/// The shader squares radius_px and emits a binary 0/1 edge alpha.
+/// V38 RETAIL-POLISH: keep the proven V34/V36/V37 JFA and refine only
+/// the final alpha profile. Still only two texture reads.
+/// c0 = (screen_width, screen_height, core_radius_px, halo_radius_px).
 const jfa_decode_src =
     "ps_3_0\n" ++
-    "def c1, 0.0, 1.0, -0.002, 0.0\n" ++
+    // c1 = zero, one, halo alpha, -interior threshold
+    "def c1, 0.0, 1.0, 0.17, -0.002\n" ++
+    // pale Retail-like yellow-green
+    "def c2, 1.0, 0.96, 0.62, 0.0\n" ++
+    // core alpha
+    "def c3, 0.84, 0.0, 0.0, 0.0\n" ++
     "dcl_2d s0\n" ++
     "dcl_2d s1\n" ++
     "dcl_texcoord0 v0\n" ++
-    // Nearest seed and pixel-space squared distance.
+    // Nearest JFA seed and pixel-space squared distance.
     "texld r0, v0, s0\n" ++
     "sub r1.xy, v0.xy, r0.xy\n" ++
     "mul r1.xy, r1.xy, c0.xy\n" ++
-    "dp2add r1.z, r1, r1, c0.w\n" ++
-    // Seed colour.
-    "texld r2, r0, s1\n" ++
-    // Hard radius test: dist² < radius² => alpha 1, otherwise 0.
-    "mov r3.x, c0.z\n" ++
-    "mul r3.x, r3.x, r3.x\n" ++
-    "sub r3.y, r1.z, r3.x\n" ++
-    "mov r6.w, c1.y\n" ++
-    "cmp r4.w, r3.y, c0.w, r6.w\n" ++
-    // Do not paint over the model interior.
+    "dp2add r1.z, r1, r1, c1.x\n" ++
+    // Squared core and halo radii.
+    "mul r3.x, c0.z, c0.z\n" ++
+    "mul r3.y, c0.w, c0.w\n" ++
+    // Smooth halo factor = saturate((halo² - dist²) / (halo² - core²)).
+    "sub r3.z, r3.y, r3.x\n" ++
+    "rcp r3.z, r3.z\n" ++
+    "sub r4.x, r3.y, r1.z\n" ++
+    "mul r4.x, r4.x, r3.z\n" ++
+    "max r4.x, r4.x, c1.x\n" ++
+    "min r4.x, r4.x, c1.y\n" ++
+    "mul r6.w, r4.x, c1.z\n" ++
+    // Solid-but-soft core inside the inner radius.
+    "sub r4.y, r1.z, r3.x\n" ++
+    "cmp r6.w, r4.y, r6.w, c3.x\n" ++
+    // Never paint over the original model interior.
     "texld r5, v0, s1\n" ++
-    "add r5.x, r5.a, c1.z\n" ++
-    "cmp r4.w, r5.x, c0.w, r4.w\n" ++
-    "mov r4.xyz, r2.xyz\n" ++
-    "mov oC0, r4\n";
+    "add r5.x, r5.a, c1.w\n" ++
+    "cmp r6.w, r5.x, c1.x, r6.w\n" ++
+    "mov r6.xyz, c2.xyz\n" ++
+    "mov oC0, r6\n";
 
 /// Debug: composite silhouette RT directly. Forces alpha to 1.0 where silhouette
 /// has any content (alpha >= 0.002), 0.0 elsewhere. Bypasses JFA entirely.
@@ -1417,8 +1437,8 @@ fn runJfaPipeline(device: *anyopaque) void {
         if (saved_rt0) |rt| deviceSetRenderTarget(device, 0, rt);
         deviceSetTexture(device, 0, rt_jfa_a_tex);
         deviceSetTexture(device, 1, rt_silhouette_tex);
-        // V34: hard 3px edge unchanged; JFA seed precision is now full-float.
-        c0 = [4]f32{ fw, fh, 3.0, 0.0 };
+        // V38: slightly stronger 1.5px core with a subtler 3.1px falloff.
+        c0 = [4]f32{ fw, fh, 1.50, 3.10 };
         deviceSetPSConstF(device, 0, &c0);
         deviceSetPtr(device, types.VT.SetPixelShader, jfa_decode_ps.?);
         deviceSetRS(device, types.D3DRS.ALPHABLENDENABLE, 1);
@@ -1626,6 +1646,7 @@ fn getD3D9VTable() ?[*]usize {
 
 pub fn installHooks() bool {
     if (hooks_installed) return true;
+    if (shutdown_started) return false;
 
     const vtable_ptr = getD3D9VTable() orelse return false;
     d3d9_vtable = vtable_ptr;

src/outline/outline.zig

@@ -109,8 +109,11 @@ pub fn initD3D9Deferred() void {
 /// Remove all outline hooks. Called during DLL_PROCESS_DETACH.
 pub fn cleanup() void {
     if (g_is_hook_owner) {
-        d3d9_hook.removeHooks();
+        // ExitFix: stop any future self-heal first, then detach the model hooks
+        // that can call lateRehookIfLost(), and only then tear down D3D9.
+        d3d9_hook.beginShutdown();
         model_hook.removeHooks();
+        d3d9_hook.removeHooks();
         log.close();
         mod_mutex.release(&g_mutex);
     }