diff --git a/.agents/skills/optimizing-rust-performance/SKILL.md b/.agents/skills/optimizing-rust-performance/SKILL.md
new file mode 100644
index 0000000..5baf775
--- /dev/null
+++ b/.agents/skills/optimizing-rust-performance/SKILL.md
@@ -0,0 +1,210 @@
+---
+name: optimizing-rust-performance
+description: |
+  审查、重构或编写 Rust 代码时主动识别性能瓶颈并应用优化模式。触发关键词：
+  "optimize"、"perf"、"性能优化"、"加速"、"hot path"、"热点路径"、
+  "make it faster"、"reduce allocation"、"zero-copy"、"零拷贝"、
+  以及任何对 .rs 文件的 review/refactor 请求、或代码中出现 Vec::remove /
+  clone / Vec<String> 在循环中 / String::from / filter 后 collect 等可疑模式时。
+allowed-tools:
+  - Read
+  - Edit
+  - Grep
+  - Glob
+metadata:
+  trigger: Rust 代码性能审查 / 重构 / 热点路径优化 / 减少 heap 分配
+  source: 基于 Rust 性能优化通用最佳实践 + 真实 baseline 测试提炼
+---
+
+# Rust 性能优化（Performance Optimization）
+
+编写或审查 Rust 代码时，**主动**识别性能瓶颈并应用优化模式。核心原则：
+**按固定优先级排序**，**判断触发条件即应用**，**量化收益**，**用 profiling 验证**。
+
+## 优化心智模型（必须按此顺序判断）
+
+```
+优化不是"想到什么改什么"。永远按此优先级评估：
+
+1. 算法与复杂度  — O(n)→O(1)、O(n²)→O(n log n)。最大收益，优先看。
+2. 内存分配     — 减少 heap 分配，栈优先，能零拷贝就零拷贝。
+3. 数据布局     — cache locality、字段顺序、对齐、SoA vs AoS。
+4. 并发         — 减少锁竞争，考虑 lock-free / 无锁结构。
+
+低层级优化在高层级问题存在时收益微乎其微。先看复杂度，再看分配。
+```
+
+## 铁律：判断到触发条件就应用，不要只"提及"
+
+> **发现反模式 → 直接改。不要说"可以考虑用 X"然后不改。**
+>
+> 如果触发条件命中，你必须（a）在改进后的代码里实际应用该模式，
+> （b）说明改了什么、为什么快。把"应该用 Cow"挂在嘴边却不写进代码，
+> 等于没优化。
+
+## 核心模式（触发条件 → 动作）
+
+### 模式 1：集合删除（顺序无关时）
+
+**触发**：代码用 `.remove(index)` 删除 `Vec` 元素，且调用方不关心顺序。
+
+**动作**：改用 `.swap_remove(index)` —— O(1)，把末尾元素换到被删位置，无内存搬移。
+
+```rust
+// ❌ O(n)：删中间元素要把后面所有元素左移
+self.items.remove(index);
+
+// ✅ O(1)：末尾元素换位，不保证顺序
+self.items.swap_remove(index);
+```
+
+**收益**：O(n) → O(1) 的内存搬移。
+
+### 模式 2：集合过滤
+
+**触发**：代码用 `for` 循环 + 条件 `push` 到新 `Vec`，或 `filter()` 后 `collect()` 再赋值。
+
+**动作**：用 `.retain()` / `.retain_mut()` 原地 O(n) 过滤，避免第二个 `Vec` 分配和逐元素 `clone`。
+
+```rust
+// ❌ 分配第二个 Vec + 每个元素 clone
+let mut kept = Vec::new();
+for item in &self.items {
+    if item.in_stock { kept.push(item.clone()); }
+}
+self.items = kept;
+
+// ✅ 原地过滤，零额外分配，零 clone
+self.items.retain(|item| item.in_stock);
+```
+
+**收益**：省一次堆分配 + N 次 clone。
+
+### 模式 3：所有权转移，避免 clone
+
+**触发**：从 `&mut T` / `Option<T>` 取值时用了 `.clone()`，或想"取出旧值替换为默认"。
+
+**动作**：
+
+- `Option<T>` 取值 → `option.take()`（取出并留 `None`，无 clone）
+- `T: Default` 取出旧值 → `std::mem::take(dest)`（旧值返回，`dest` 变默认）
+- 交换值 → `std::mem::replace(dest, src)`（无深拷贝）
+
+```rust
+// ❌ clone 后原值仍在，字段没被清空（还可能是 bug）
+fn clear_promo(&mut self) -> Option<PromoCode> {
+    self.active_promo.clone()
+}
+
+// ✅ take：取出所有权，字段变 None，无 clone
+fn clear_promo(&mut self) -> Option<PromoCode> {
+    self.active_promo.take()
+}
+```
+
+**收益**：消除一次堆分配（clone 的 `String`/`Vec` 等）。
+
+### 模式 4：栈优先（短生命周期小集合）
+
+**触发**：循环内频繁分配小而短命的 `Vec`/`String`（如"通常 2-3 个元素"的辅助返回值）。
+
+**动作**：用 `SmallVec`/`TinyVec` 把小负载（<4 或 <8 元素）放栈上，溢出才上堆。
+
+```rust
+// ❌ 每次 tags_for 都堆分配一个 Vec（通常只有 2-3 个 tag）
+fn tags_for(&self, id: u32) -> Vec<String> {
+    vec![format!("tag-{id}-a"), format!("tag-{id}-b")]
+}
+
+// ✅ SmallVec：2-3 个元素常驻栈，无堆分配
+fn tags_for(&self, id: u32) -> SmallVec<[String; 4]> {
+    smallvec![format!("tag-{id}-a"), format!("tag-{id}-b")]
+}
+```
+
+**收益**：栈分配 vs 堆分配——省掉 malloc/free 和可能的 cache miss。
+
+### 模式 5：Copy-on-Write 延迟分配
+
+**触发**：字符串/切片处理大多只读，偶尔才需要修改或拥有所有权；返回类型是 `String` 但其实常无需分配。
+
+**动作**：用 `std::borrow::Cow` 封装——只读时零分配借用，真正写时才 clone。
+
+```rust
+use std::borrow::Cow;
+
+// ❌ 永远堆分配，即使输入已经全是小写无需改动
+fn normalize(name: &str) -> String {
+    name.trim().to_lowercase()
+}
+
+// ✅ 只在确实需要改写（trim 砍掉字符 / 含大写）时才分配；
+//    纯小写无空白的输入零分配，原样借用返回
+fn normalize<'a>(name: &'a str) -> Cow<'a, str> {
+    let trimmed = name.trim();
+    let needs_lower = trimmed.chars().any(|c| c.is_ascii_uppercase());
+    let needs_trim = trimmed.len() != name.len();
+    if !needs_lower && !needs_trim {
+        Cow::Borrowed(trimmed)
+    } else {
+        Cow::Owned(trimmed.to_lowercase())
+    }
+}
+```
+
+**收益**：读路径零分配；分配延迟到真正写时才发生。
+
+## 热点路径额外模式
+
+热点路径（每秒百万次调用的解析器/词法器/序列化）适用更激进的优化：
+
+- **切片代替逐字符 collect**：从原 `&str`/`&[u8]` 用范围切片 `&input[start..pos]` 取词素，而不是把 `char` push 进 `Vec<char>` 再 `collect::<String>()`（省双重分配 + 双重拷贝）。
+- **借用切片匹配后再拥有化**：先在借用的 `&str` 上 match 关键字（零分配），仅非关键字才 `.to_string()`。
+- **ASCII 谓词代替 Unicode 谓词**：`is_ascii_digit()` / `is_ascii_alphabetic()` 代替 `is_digit(10)` / `is_alphanumeric()`（省 Unicode 表查找）——前提是你确实只需 ASCII。
+
+## 回应规范（应用优化时必须做到）
+
+1. **先分析瓶颈**：说明当前问题（"这会导致 O(n) 内存搬移" / "这触发一次堆分配"）。
+2. **给出优化代码**：干净、生产可用的 Rust，实际应用模式（不只是提及）。
+3. **量化理论收益**：说明*为什么*更快（堆 vs 栈、O(n) vs O(1)、cache locality）。
+4. **提示 profiling**：提醒用 `Criterion`（微基准）或 `Flamegraph` 验证，不要盲信理论。
+
+**量化收益示例**（写到改动说明里）：
+
+| 改动                       | 复杂度 / 分配变化           |
+| -------------------------- | --------------------------- |
+| `remove` → `swap_remove`   | O(n) → O(1) 内存搬移        |
+| 循环+clone → `retain`      | 省 1 次堆分配 + N 次 clone  |
+| `clone` → `take`           | 省 1 次堆分配               |
+| `Vec` → `SmallVec<[T; 4]>` | 堆分配 → 栈分配（小负载时） |
+| `String` → `Cow<str>`      | 读路径：1 次分配 → 0 次     |
+
+## 自检清单（提交 Rust 改动前逐条过）
+
+- [ ] 任何 `Vec::remove` 在顺序无关处是否已换 `swap_remove`？
+- [ ] 过滤操作是否用了 `retain` / `retain_mut` 而非 collect？
+- [ ] 取值是否用了 `take` / `mem::take` / `mem::replace` 而非 `clone`？
+- [ ] 循环内小而短命的 `Vec`/`String` 是否考虑过 `SmallVec`？
+- [ ] 大多只读、偶尔写的字符串/切片是否考虑过 `Cow`？
+- [ ] 热点路径是否切片取词素而非逐字符 collect？
+- [ ] 每条改动是否说明了复杂度/分配收益，并提示 profiling 验证？
+- [ ] 引入新依赖（`smallvec` 等）或改动公开返回类型前，确认收益配得上成本。
+
+## 注意（应用前必读 —— 防止过度优化）
+
+优化有成本。应用下列模式前先权衡，**别为了用模式而用**：
+
+- **先正确再优化**：`#[inline]`、手动 SIMD、`unsafe` 等通常不是首选；先确认算法复杂度和分配已最优。
+- **新增依赖有代价**：`SmallVec`/`TinyVec` 要加 crate 依赖、增加编译时间。只为"通常 2-3 个元素"就在非热点代码里引入依赖，多半不划算——优先考虑能否直接用数组 `[T; N]` / slice / 返回迭代器。
+- **Cow/SmallVec 改返回类型是 API 传染**：把 `-> String` 改成 `-> Cow<'_, str>` 会把生命周期参数传染给所有调用方；把 `Token` 改成 `Token<'src>` 是全 crate 的 API 变更。改公开签名前确认收益配得上波及面；内部 `fn` 则无所谓。
+- **量入为出**：Cow/SmallVec 引入复杂度，确认该路径真的是热点或高频才上。非热点的小函数直接 `Vec`/`String` 更清晰。
+- **profile 验证**：理论收益不等于实测收益。用 `Criterion` 做微基准，`Flamegraph` 找真热点，别盲改。
+
+**决策速查**：
+
+| 情况                 | 推荐                                        |
+| -------------------- | ------------------------------------------- |
+| 非热点、简单工具函数 | 直接 `Vec`/`String`/`clone`，别上模式       |
+| 公开 API 返回类型    | 慎用 `Cow<'_, str>`/生命周期——会传染调用方  |
+| 真热点 + 内部函数    | 放心上 Cow/SmallVec/借用切片                |
+| 新增依赖             | 先看能否用 `[T; N]` / 迭代器 / 现有类型替代 |
diff --git a/assets/shaders/black_hole.wgsl b/assets/shaders/black_hole.wgsl
index 0433c96..37b7be8 100644
--- a/assets/shaders/black_hole.wgsl
+++ b/assets/shaders/black_hole.wgsl
@@ -39,6 +39,15 @@ struct BlackHoleUniforms {
     bloom_strength: f32,
     exposure: f32,
     _pad5: f32,
+    disk_half_thickness: f32,
+    filament_freq: f32,
+    filament_sharpness: f32,
+    density_freq: f32,
+    density_strength: f32,
+    arm_count: f32,
+    arm_tightness: f32,
+    arm_strength: f32,
+    disk_quality: u32,
 };
 
 @group(#{MATERIAL_BIND_GROUP}) @binding(0) var<uniform> uniforms: BlackHoleUniforms;
@@ -206,10 +215,15 @@ fn value_noise3(p: vec3<f32>) -> f32 {
 }
 
 fn fbm3(p: vec3<f32>, octaves: u32) -> f32 {
+    // MAX_OCTAVES-with-break: the conservative WebGPU form for a runtime-
+    // chosen octave count. Older drivers can miscompile non-constant loop
+    // bounds; this branch is the first to pass dynamic octaves here.
+    const MAX_OCTAVES = 6u;
     var sum = 0.0;
     var amp = 0.5;
     var freq = 1.0;
-    for (var i: u32 = 0u; i < octaves; i = i + 1u) {
+    for (var i: u32 = 0u; i < MAX_OCTAVES; i = i + 1u) {
+        if (i >= octaves) { break; }
         sum = sum + amp * value_noise3(p * freq);
         freq = freq * 2.0;
         amp = amp * 0.5;
@@ -217,16 +231,75 @@ fn fbm3(p: vec3<f32>, octaves: u32) -> f32 {
     return sum;
 }
 
+// Ridged multifractal noise: 1 - |2n-1| turns value-noise gradients into
+// sharp ridges (peak where n=0.5, zero at n=0 and n=1). Raising to
+// `sharpness` thins the ridges into filaments. MAX_OCTAVES-with-break is
+// the conservative WebGPU form for a runtime-chosen octave count.
+fn ridged_fbm(p: vec3<f32>, octaves: u32, sharpness: f32) -> f32 {
+    const MAX_OCTAVES = 6u;
+    var sum = 0.0;
+    var amp = 0.5;
+    var freq = 1.0;
+    for (var i: u32 = 0u; i < MAX_OCTAVES; i = i + 1u) {
+        if (i >= octaves) { break; }
+        let n = value_noise3(p * freq);
+        let ridge = 1.0 - abs(2.0 * n - 1.0);
+        sum = sum + amp * pow(ridge, sharpness);
+        freq = freq * 2.0;
+        amp = amp * 0.5;
+    }
+    return sum;
+}
+
 fn disk_noise(pos: vec3<f32>, t: f32) -> f32 {
     let warp = fbm3(pos * 0.8 + vec3<f32>(0.0, 0.0, t * 0.1), 3u);
     let n = fbm3(pos * 2.0 + warp * 1.5 + vec3<f32>(0.0, 0.0, t * 0.3), 4u);
     return n;
 }
 
-fn disk_color(pos: vec3<f32>, dir: vec3<f32>) -> vec3<f32> {
-    let r = length(vec2<f32>(pos.x, pos.z));
-    let phi = atan2(pos.z, pos.x);
+// Result of a disk color query: emitted radiance + opacity contribution.
+// Both the volumetric and flat paths return this struct so the main loop
+// can treat them uniformly.
+struct DiskSample {
+    color: vec3<f32>,
+    density: f32,
+}
 
+// Radial temperature gradient: white-hot inner → deep-orange outer.
+fn temperature_color(t: f32) -> vec3<f32> {
+    return mix(vec3<f32>(1.0, 0.95, 0.85), vec3<f32>(1.0, 0.45, 0.12), clamp(t, 0.0, 1.0));
+}
+
+// Radial brightness falloff (∝ 1/r² from the inner edge).
+fn radial_falloff(r: f32, inner: f32) -> f32 {
+    return 1.0 / pow(r / inner, 2.0);
+}
+
+// Cylindrical radius in the disk plane.
+fn r_of(pos: vec3<f32>) -> f32 {
+    return length(vec2<f32>(pos.x, pos.z));
+}
+
+// Relativistic Doppler beaming. `dir` is the ray direction (disk-local).
+fn apply_doppler(col: vec3<f32>, pos: vec3<f32>, dir: vec3<f32>) -> vec3<f32> {
+    let phi = atan2(pos.z, pos.x);
+    let v_orbital = sqrt(uniforms.rs / (2.0 * r_of(pos)));
+    let tangent = normalize(vec3<f32>(-sin(phi), 0.0, cos(phi)));
+    let vdotn = dot(tangent * v_orbital, -dir);
+    let gamma = 1.0 / sqrt(max(1.0 - v_orbital * v_orbital, 1e-4));
+    if (uniforms.doppler_enabled == 0u) {
+        return col;
+    }
+    let delta = 1.0 / (gamma * (1.0 - vdotn));
+    let doppler = pow(delta, 3.0) * uniforms.doppler_strength;
+    return col * doppler;
+}
+
+// Off-tier fallback: zero-thickness disk, single sample, fixed alpha.
+// Preserves the exact pre-volumetric appearance. Returns DiskSample so the
+// main loop dispatches both paths uniformly.
+fn disk_color_flat(pos: vec3<f32>, dir: vec3<f32>) -> DiskSample {
+    let r = r_of(pos);
     let rot = uniforms.time * uniforms.disk_rotation_speed / pow(r, 1.5);
     // Domain-warped FBM for feathered/smoky gas texture. The Keplerian shear
     // (rot ∝ 1/r^1.5) is folded into the noise flow term so inner radii flow
@@ -234,24 +307,65 @@ fn disk_color(pos: vec3<f32>, dir: vec3<f32>) -> vec3<f32> {
     let noise = disk_noise(vec3<f32>(pos.x * 0.3, pos.z * 0.3, rot), uniforms.time);
 
     let t = (r - uniforms.disk_inner) / (uniforms.disk_outer - uniforms.disk_inner);
-    let tcol = mix(vec3<f32>(1.0, 0.95, 0.85), vec3<f32>(1.0, 0.45, 0.12), clamp(t, 0.0, 1.0));
+    let tcol = temperature_color(t);
+    let falloff = radial_falloff(r, uniforms.disk_inner);
 
-    let falloff = 1.0 / pow(r / uniforms.disk_inner, 2.0);
+    var col = tcol * (0.6 + 0.4 * noise) * falloff * uniforms.disk_brightness;
+    col = apply_doppler(col, pos, dir);
 
-    var col = tcol * (0.6 + 0.4 * noise) * falloff;
+    return DiskSample(vec3<f32>(col), 0.85);
+}
 
-    let v_orbital = sqrt(uniforms.rs / (2.0 * r));
-    let tangent = normalize(vec3<f32>(-sin(phi), 0.0, cos(phi)));
-    let vdotn = dot(tangent * v_orbital, -dir);
-    let gamma = 1.0 / sqrt(max(1.0 - v_orbital * v_orbital, 1e-4));
-    var doppler = 1.0;
-    if (uniforms.doppler_enabled != 0u) {
-        let delta = 1.0 / (gamma * (1.0 - vdotn));
-        doppler = pow(delta, 3.0) * uniforms.doppler_strength;
-    }
-    col *= doppler;
+// Volumetric disk color. Noise is sampled in POLAR coordinates (r_norm,
+// phi·freq, height) so turbulence flows tangentially — the correct pattern
+// for a rotating fluid. Three layers multiply: ridged filaments (brightness),
+// soft density clumping, logarithmic-spiral arms advected by Keplerian shear.
+fn disk_color_volumetric(pos: vec3<f32>, dir: vec3<f32>) -> DiskSample {
+    let r = r_of(pos);
+    let phi = atan2(pos.z, pos.x);
+    let rot = uniforms.time * uniforms.disk_rotation_speed / pow(r, 1.5);
 
-    return col * uniforms.disk_brightness;
+    // Octave triplet from the quality tier.
+    let q = uniforms.disk_quality;
+    var filament_octaves = 5u; var density_octaves = 4u; var warp_octaves = 3u;
+    if (q == 1u) { filament_octaves = 3u; density_octaves = 2u; warp_octaves = 2u; }
+    else if (q == 2u) { filament_octaves = 4u; density_octaves = 3u; warp_octaves = 3u; }
+    // q == 3u keeps the High defaults above; q == 0u is never passed here.
+
+    // Polar sample coordinate: (r normalized, angle × freq + Keplerian flow,
+    // height within slab). The flow term advects the noise so inner radii
+    // (faster rotation) drift ahead of outer radii — differential rotation.
+    let r_norm = r / uniforms.disk_inner;
+    let h = pos.y / max(uniforms.disk_half_thickness, 1e-3);
+    let sp = vec3<f32>(r_norm, phi * 2.5 + rot, h);
+
+    // Domain warp in polar space: distorts sample coords so filaments bend.
+    let warp = fbm3(sp * 0.8, warp_octaves);
+
+    // Layer 1: ridged bright filaments (polar-sampled → tangential streaks).
+    let filament = ridged_fbm(sp * uniforms.filament_freq + warp * 1.5,
+                              filament_octaves, uniforms.filament_sharpness);
+
+    // Layer 2: density clumping (soft ramp, no hard cut → avoids patchiness).
+    let density_noise = fbm3(sp * uniforms.density_freq + warp, density_octaves);
+    let base_density = (0.35 + 0.65 * density_noise) * uniforms.density_strength;
+
+    // Layer 3: logarithmic-spiral arm modulation, advected by Keplerian shear.
+    let arm_phase = phi * uniforms.arm_count + log(max(r, 0.1)) * uniforms.arm_tightness - rot;
+    let arm = 0.5 + 0.5 * cos(arm_phase);
+    let arm_mod = mix(1.0, pow(arm, 2.0), uniforms.arm_strength);
+
+    let total_density = base_density * arm_mod;
+    let brightness = 0.5 + filament;
+
+    let t = (r - uniforms.disk_inner) / (uniforms.disk_outer - uniforms.disk_inner);
+    let tcol = temperature_color(t);
+    let falloff = radial_falloff(r, uniforms.disk_inner);
+
+    var col = tcol * brightness * falloff * uniforms.disk_brightness;
+    col = apply_doppler(col, pos, dir);
+
+    return DiskSample(vec3<f32>(col), total_density);
 }
 
 // --- planets ---
@@ -444,13 +558,52 @@ fn fragment(in: VertexOutput) -> @location(0) vec4<f32> {
             break;
         }
 
-        if (disk_hit(prev, new_pos)) {
-            let ty = prev.y / (prev.y - new_pos.y);
-            let hit = mix(prev, new_pos, vec3<f32>(ty));
-            let dc = disk_color(hit, new_dir);
-            let a = 0.85;
-            accum_color += (1.0 - accum_alpha) * dc * a;
-            accum_alpha += (1.0 - accum_alpha) * a;
+        // --- volumetric disk ---
+        if (uniforms.disk_quality == 0u) {
+            // Off tier: zero-thickness single midplane sample, fixed alpha.
+            if (disk_hit(prev, new_pos)) {
+                let ty = prev.y / (prev.y - new_pos.y);
+                let hit = mix(prev, new_pos, vec3<f32>(ty));
+                let s = disk_color_flat(hit, new_dir);
+                accum_color += (1.0 - accum_alpha) * s.color * s.density;
+                accum_alpha += (1.0 - accum_alpha) * s.density;
+                if (accum_alpha > 0.99) { break; }
+            }
+        } else {
+            // Volumetric tier. Single unified path: clip THIS step's segment
+            // to the disk slab |y|<=half_thickness, sample once at the clipped
+            // segment's midpoint, accumulate × the in-slab segment length.
+            // The previous design used two overlapping paths (in-slab step +
+            // at-plane edge-capture) with inconsistent angle-dependent weights
+            // (step_len vs thickness_proj) that produced radial spokes.
+            let H = uniforms.disk_half_thickness;
+            let dy = new_pos.y - prev.y;
+            var seg_t0 = 0.0;
+            var seg_t1 = 1.0;
+            var has_slab_seg = false;
+            if (abs(dy) < 1e-6) {
+                // Step is parallel to the slab plane: in-slab iff prev is in-slab.
+                has_slab_seg = abs(prev.y) <= H;
+            } else {
+                // Clip the parametric line prev+t*(new_pos-prev) to |y|<=H.
+                let ta = (H - prev.y) / dy;
+                let tb = (-H - prev.y) / dy;
+                seg_t0 = clamp(min(ta, tb), 0.0, 1.0);
+                seg_t1 = clamp(max(ta, tb), 0.0, 1.0);
+                has_slab_seg = seg_t1 > seg_t0;
+            }
+
+            if (has_slab_seg) {
+                let mid_t = (seg_t0 + seg_t1) * 0.5;
+                let mid = mix(prev, new_pos, vec3<f32>(mid_t));
+                let mid_r = r_of(mid);
+                if (mid_r >= uniforms.disk_inner && mid_r <= uniforms.disk_outer) {
+                    let s = disk_color_volumetric(mid, new_dir);
+                    let seg_len = (seg_t1 - seg_t0) * length(new_pos - prev);
+                    accum_color += (1.0 - accum_alpha) * s.color * s.density * seg_len;
+                    accum_alpha += (1.0 - accum_alpha) * s.density * seg_len;
+                }
+            }
             if (accum_alpha > 0.99) { break; }
         }
 
diff --git a/skills-lock.json b/skills-lock.json
index 6903533..54754f2 100644
--- a/skills-lock.json
+++ b/skills-lock.json
@@ -67,6 +67,12 @@
       "skillPath": "skills/engineering/improve-codebase-architecture/SKILL.md",
       "computedHash": "8bf292143ca93b00276a0de0fc5f84f2381f4690c5b0d32e99fa283a072f392f"
     },
+    "optimizing-rust-performance": {
+      "source": "DefectingCat/optimizing-rust-performance",
+      "sourceType": "github",
+      "skillPath": "SKILL.md",
+      "computedHash": "96bf0221d116e617055f5dc038bb33d92dd61b7808d804f92da78d32c01e6934"
+    },
     "prototype": {
       "source": "mattpocock/skills",
       "sourceType": "github",
diff --git a/src/camera.rs b/src/camera.rs
index 34cf86a..3b0a780 100644
--- a/src/camera.rs
+++ b/src/camera.rs
@@ -16,11 +16,12 @@ pub struct OrbitCamera {
 impl Default for OrbitCamera {
     fn default() -> Self {
         Self {
-            yaw: 0.0,
-            // ~0.7 rad (40°) is a comfortable 3/4 view of the disk. The basis()
-            // no longer has a gimbal pole, so any pitch is safe; this is just a
-            // nice default angle, not a pole-avoidance choice.
-            pitch: 0.7,
+            yaw: -1.065,
+            // ~0.335 rad (19°) below the disk plane. The lensed far side of the
+            // disk then arcs over the top of the shadow — the iconic framing. The
+            // basis() has no gimbal pole, so any pitch is safe; this is just a nice
+            // default angle, not a pole-avoidance choice.
+            pitch: -0.335,
             distance: 30.0,
             fov: 1.0,         // radians
         }
diff --git a/src/params.rs b/src/params.rs
index 337063f..812c913 100644
--- a/src/params.rs
+++ b/src/params.rs
@@ -21,6 +21,41 @@ impl BloomQuality {
     }
 }
 
+/// Disk volumetric rendering quality. Gates noise octave counts.
+#[allow(dead_code)] // consumed starting Task 7 (tier dispatch) + Task 8 (egui)
+#[derive(Clone, Copy, PartialEq, Eq, Default, Debug)]
+pub enum DiskQuality {
+    Off,     // flat zero-thickness fallback (current appearance)
+    Low,     // 3/2/2 octaves — web default
+    Medium,  // 4/3/3 octaves
+    #[default]
+    High,    // 5/4/3 octaves — desktop default
+}
+
+impl DiskQuality {
+    /// Returns (filament_octaves, density_octaves, warp_octaves).
+    /// Off returns zeros; the shader dispatches to the flat path instead.
+    #[allow(dead_code)] // read in Task 7's tier dispatch
+    pub fn octaves(self) -> (u32, u32, u32) {
+        match self {
+            DiskQuality::Off => (0, 0, 0),
+            DiskQuality::Low => (3, 2, 2),
+            DiskQuality::Medium => (4, 3, 3),
+            DiskQuality::High => (5, 4, 3),
+        }
+    }
+
+    /// Tier as a u32 for the WGSL uniform selector.
+    pub fn as_u32(self) -> u32 {
+        match self {
+            DiskQuality::Off => 0,
+            DiskQuality::Low => 1,
+            DiskQuality::Medium => 2,
+            DiskQuality::High => 3,
+        }
+    }
+}
+
 /// All tunable black-hole parameters. Edited by the egui panel (Task 17),
 /// mirrored into BlackHoleUniforms each frame (Task 7).
 #[derive(Resource, Clone)]
@@ -54,6 +89,16 @@ pub struct BlackHoleParams {
     pub bloom_strength: f32,
     pub exposure: f32,
     pub bloom_quality: BloomQuality,
+    // Disk turbulence (Phase 3.1: volumetric disk)
+    pub disk_half_thickness: f32,
+    pub filament_freq: f32,
+    pub filament_sharpness: f32,
+    pub density_freq: f32,
+    pub density_strength: f32,
+    pub arm_count: f32,
+    pub arm_tightness: f32,
+    pub arm_strength: f32,
+    pub disk_quality: DiskQuality,
 }
 
 impl Default for BlackHoleParams {
@@ -80,6 +125,15 @@ impl Default for BlackHoleParams {
             bloom_strength: 0.8,
             exposure: 1.0,
             bloom_quality: if cfg!(target_arch = "wasm32") { BloomQuality::Low } else { BloomQuality::High },
+            disk_half_thickness: if cfg!(target_arch = "wasm32") { 0.2 } else { 0.3 },
+            filament_freq: 1.0,
+            filament_sharpness: 2.0,
+            density_freq: 0.8,
+            density_strength: 1.0,
+            arm_count: 2.0,
+            arm_tightness: 2.0,
+            arm_strength: 0.5,
+            disk_quality: if cfg!(target_arch = "wasm32") { DiskQuality::Low } else { DiskQuality::High },
         }
     }
 }
diff --git a/src/render/material.rs b/src/render/material.rs
index 2a65bef..12a74b8 100644
--- a/src/render/material.rs
+++ b/src/render/material.rs
@@ -43,6 +43,16 @@ pub struct BlackHoleUniforms {
     pub bloom_strength: f32,
     pub exposure: f32,
     pub _pad5: f32,
+    // Disk volumetric (Phase 3.1)
+    pub disk_half_thickness: f32,
+    pub filament_freq: f32,
+    pub filament_sharpness: f32,
+    pub density_freq: f32,
+    pub density_strength: f32,
+    pub arm_count: f32,
+    pub arm_tightness: f32,
+    pub arm_strength: f32,
+    pub disk_quality: u32,
 }
 
 impl Default for BlackHoleUniforms {
@@ -79,6 +89,15 @@ impl Default for BlackHoleUniforms {
             bloom_strength: 0.8,
             exposure: 1.0,
             _pad5: 0.0,
+            disk_half_thickness: 0.3,
+            filament_freq: 1.0,
+            filament_sharpness: 2.0,
+            density_freq: 0.8,
+            density_strength: 1.0,
+            arm_count: 2.0,
+            arm_tightness: 2.0,
+            arm_strength: 0.5,
+            disk_quality: 3, // High
         }
     }
 }
diff --git a/src/render/plugin.rs b/src/render/plugin.rs
index f83284f..7c91d6f 100644
--- a/src/render/plugin.rs
+++ b/src/render/plugin.rs
@@ -620,6 +620,15 @@ fn mirror_params(
         u.bloom_threshold = params.bloom_threshold;
         u.bloom_strength = params.bloom_strength;
         u.exposure = params.exposure;
+        u.disk_half_thickness = params.disk_half_thickness;
+        u.filament_freq = params.filament_freq;
+        u.filament_sharpness = params.filament_sharpness;
+        u.density_freq = params.density_freq;
+        u.density_strength = params.density_strength;
+        u.arm_count = params.arm_count;
+        u.arm_tightness = params.arm_tightness;
+        u.arm_strength = params.arm_strength;
+        u.disk_quality = params.disk_quality.as_u32();
     }
     // Update brightpass threshold (live-tunable).
     for (_, mat) in brightpass_materials.iter_mut() {
diff --git a/src/ui.rs b/src/ui.rs
index 522a20d..1ef1ba9 100644
--- a/src/ui.rs
+++ b/src/ui.rs
@@ -36,6 +36,30 @@ pub fn ui_system(
                         ui.add(egui::Slider::new(&mut params.disk_brightness, 0.0..=3.0).text("Brightness"));
                         ui.add(egui::Slider::new(&mut params.disk_rotation_speed, 0.0..=3.0).text("Rotation speed"));
                     });
+                egui::CollapsingHeader::new("Disk Turbulence")
+                    .default_open(true)
+                    .show(ui, |ui| {
+                        use crate::params::DiskQuality;
+                        let mut q = params.disk_quality;
+                        egui::ComboBox::from_label("Disk quality")
+                            .selected_text(format!("{:?}", q))
+                            .show_ui(ui, |ui| {
+                                ui.selectable_value(&mut q, DiskQuality::Off, "Off");
+                                ui.selectable_value(&mut q, DiskQuality::Low, "Low");
+                                ui.selectable_value(&mut q, DiskQuality::Medium, "Medium");
+                                ui.selectable_value(&mut q, DiskQuality::High, "High");
+                            });
+                        params.disk_quality = q;
+                        let on = q != DiskQuality::Off;
+                        ui.add_enabled(on, egui::Slider::new(&mut params.disk_half_thickness, 0.05..=1.0).text("Half thickness"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.filament_freq, 0.2..=4.0).text("Filament frequency"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.filament_sharpness, 1.0..=6.0).text("Filament sharpness"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.density_freq, 0.2..=3.0).text("Density frequency"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.density_strength, 0.0..=2.0).text("Density strength"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.arm_count, 0.0..=6.0).text("Arm count"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.arm_tightness, 0.0..=6.0).text("Arm tightness"));
+                        ui.add_enabled(on, egui::Slider::new(&mut params.arm_strength, 0.0..=1.0).text("Arm strength"));
+                    });
                 egui::CollapsingHeader::new("Doppler").show(ui, |ui| {
                     ui.checkbox(&mut params.doppler_enabled, "Enabled");
                     ui.add_enabled(params.doppler_enabled, egui::Slider::new(&mut params.doppler_strength, 0.0..=3.0).text("Strength"));