fix(renderer): preserve float precision through bilinear un-premultiply

sample_bilinear_rgba rounded premultiplied RGB to u8 before the
un-premultiply step, losing sub-pixel precision and producing dark/
bright halos on semi-transparent sprite edges after rotation.

Keep RGB in float space through interpolation + un-premultiply and
only round/clamp once at the end.

Co-authored-by: DeepSeek V4 <deepseek@v4.ai>

Author: zachelnet

crates/koharu-app/src/renderer.rs

@@ -1126,22 +1126,28 @@ fn sample_bilinear_rgba(src: &RgbaImage, x: f32, y: f32) -> Rgba<u8> {
     let p11 = get_premul(src.get_pixel(x1 as u32, y1 as u32));
 
     let lerp = |a: f32, b: f32, t: f32| a + (b - a) * t;
-    let mut out = [0u8; 4];
-    for (i, out_ch) in out.iter_mut().enumerate().take(3) {
+
+    // Interpolate premultiplied RGB in float space (don't round yet).
+    let mut rgb_premul = [0.0f32; 3];
+    for i in 0..3 {
         let top = lerp(p00[i], p10[i], wx);
         let bottom = lerp(p01[i], p11[i], wx);
-        *out_ch = lerp(top, bottom, wy).round().clamp(0.0, 255.0) as u8;
+        rgb_premul[i] = lerp(top, bottom, wy);
     }
-    // Alpha is interpolated in straight-alpha space (premultiplied alpha = original alpha).
+
+    // Alpha is interpolated in straight-alpha space.
     let top_a = lerp(p00[3], p10[3], wx);
     let bottom_a = lerp(p01[3], p11[3], wx);
-    let alpha = lerp(top_a, bottom_a, wy).round().clamp(0.0, 255.0);
-    out[3] = alpha as u8;
+    let alpha = lerp(top_a, bottom_a, wy).clamp(0.0, 255.0);
+
+    let mut out = [0u8; 4];
+    let alpha_u8 = alpha.round() as u8;
+    out[3] = alpha_u8;
 
-    // Un-premultiply: divide R, G, B by alpha (unless fully transparent).
-    if alpha > 0.0 {
-        for out_ch in out.iter_mut().take(3) {
-            *out_ch = ((*out_ch as f32) * 255.0 / alpha).round().clamp(0.0, 255.0) as u8;
+    // Un-premultiply once, then round — preserves sub-pixel precision.
+    if alpha_u8 != 0 {
+        for i in 0..3 {
+            out[i] = (rgb_premul[i] * 255.0 / alpha).round().clamp(0.0, 255.0) as u8;
         }
     }
     Rgba(out)