Fix bloom glow flicker on static geometry under TAA jitter

TAA subpixel jitter makes rasterized HDR specular highlights oscillate
per frame. The old RGB min/max neighborhood clamp re-clamped the
converged history against jitter-driven extremes every frame, leaving a
periodic residual that the bloom threshold then amplified into visible
glow pulsing on static models.

- TAA: replace the RGB AABB clamp with YCoCg variance clipping
  (history clamped toward the stable neighborhood mean +/- 1.25 sigma,
  intersected with the min/max box as a ghosting upper bound)
- Bloom extract: 5-tap cross prefilter before the threshold spreads
  surviving 1px jitter energy over a ~4x4 footprint so the glow no
  longer pulses

shaders/taa.frag

@@ -2,7 +2,9 @@
 
 // TAA resolve：时域抗锯齿。
 //   1. velocity 反推历史像素位置，取历史累积色
-//   2. 邻域 clamp（3×3 AABB）压制鬼影：历史色超出当前帧邻域范围的部分丢弃
+//   2. 方差裁剪（YCoCg）：历史色钳向邻域均值 ± γσ（∩ AABB 极值盒）——
+//      比 RGB min/max 盒稳得多：AABB 盒子被逐帧抖动的高亮像素拽来拽去，
+//      收敛历史持续被重钳 → 周期性残差经 bloom 阈值放大 = 静态模型泛光闪烁
 //   3. mix(cur, hist, alpha) 时域累积
 // 输出写入 TAAColor，它同时也是下一帧的历史（ReadHistory 自引用）。
 
@@ -21,6 +23,19 @@ layout(push_constant) uniform PushConstants {
     float param2;
 } pc;
 
+// YCoCg 感知空间：亮度/色度分离，方差裁剪不在 HDR RGB 直接做
+//（RGB 各通道独立 σ 会偏色；Y 通道主导几何相关性）
+vec3 RgbToYCoCg(vec3 c) {
+    return vec3(0.25 * c.r + 0.5 * c.g + 0.25 * c.b,
+                0.5 * c.r - 0.5 * c.b,
+                -0.25 * c.r + 0.5 * c.g - 0.25 * c.b);
+}
+vec3 YCoCgToRgb(vec3 c) {
+    return vec3(c.x + c.y - c.z,
+                c.x + c.z,
+                c.x - c.y - c.z);
+}
+
 void main() {
     vec4 cur = texture(curImage, fragUV);
     vec2 vel = texture(velocityImage, fragUV).xy;
@@ -33,18 +48,33 @@ void main() {
         all(lessThanEqual(histUV, vec2(1.0)))) {
         vec3 hist = texture(historyImage, histUV).rgb;
 
-        // 邻域 clamp：当前帧 3×3 邻域的 min/max 包围盒
+        // 邻域统计（3×3，YCoCg）：均值 m1、均方 m2、极值盒 [mn, mx]
         vec2 ts = 1.0 / vec2(textureSize(curImage, 0));
-        vec3 minC = cur.rgb;
-        vec3 maxC = cur.rgb;
+        vec3 m1 = vec3(0.0);
+        vec3 m2 = vec3(0.0);
+        vec3 mn = vec3(1e9);
+        vec3 mx = vec3(-1e9);
         for (int y = -1; y <= 1; ++y) {
             for (int x = -1; x <= 1; ++x) {
-                vec3 c = texture(curImage, fragUV + vec2(x, y) * ts).rgb;
-                minC = min(minC, c);
-                maxC = max(maxC, c);
+                vec3 c = RgbToYCoCg(texture(curImage, fragUV + vec2(x, y) * ts).rgb);
+                m1 += c;
+                m2 += c * c;
+                mn = min(mn, c);
+                mx = max(mx, c);
             }
         }
-        vec3 clamped = clamp(hist, minC, maxC);
+        m1 /= 9.0;
+        m2 /= 9.0;
+        vec3 sigma = sqrt(max(m2 - m1 * m1, vec3(0.0)));
+
+        // 方差裁剪盒：均值 ± 1.25σ 再 ∩ 极值盒。历史被钳向"稳定均值"
+        // 而非抖动极值 → 收敛残差大幅下降（泛光闪烁的治本点）。
+        // 极值盒交集保证不引入邻域外颜色（鬼影上界）
+        const float kGamma = 1.25;
+        vec3 lo = max(m1 - kGamma * sigma, mn);
+        vec3 hi = min(m1 + kGamma * sigma, mx);
+
+        vec3 clamped = YCoCgToRgb(clamp(RgbToYCoCg(hist), lo, hi));
         result = mix(cur.rgb, clamped, pc.alpha);
     }