MilkDrop: fresh blur levels, MilkDrop's sampler wrap rules and 8-bit feedback rounding (#752)

[CaYatur]

Second quality round for #752. Three engine fixes, both checked against MilkDrop's source and measured against BeatDrop 53d83ee with the #750 patch.

Blur levels 2 and 3 were one and two frames behind

Each blur level's out texture is mipmapped, and the mipmaps were generated after the whole chain. The next level read it with texture() at a smaller size, so it got last frame's mipmap. At a preset's start, blur2 was empty on its first frame and blur3 on its first two; MilkDrop fills both. MilkDrop's blur textures have a single level and are read bilinearly. Each level now reads its source with textureLod(…, 0).

sampler_main wraps in preset shaders

MilkDrop sets a sampler's wrap mode from its name:

• a plain sampler_main is filtered and wraps (plugin.cpp:3445-3446);
• fc_ and pc_ clamp;
• bTexWrap is read only by the shader-less warp (milkdropfs.cpp:2217).

The engine followed bTexWrap in shaders too, so content leaving the screen stuck to the edge.

• Synthetic check. On a preset with bTexWrap = 0 (a warp shader shifting uv, a comp shader zooming out), MilkDrop wraps and matches sampler_fw_main exactly. The engine's difference went 3.02 → 0.15.
• Reach. 3,325 library presets (34%) have a shader and bTexWrap = 0.
• Blur reads clamped. MilkDrop's BlurPasses sets CLAMP and never resets it. Blur now always reads clamped, through a sampler object.
• Textured shapes. They are drawn after the blur, so they read clamped when a blur pass ran, as in MilkDrop.

Measured

Main → branch. Mean difference on a 16-pixel grid, median over presets without user textures or randomness.

set	f2	f10	f30	f59
150 presets with a shader and bTexWrap = 0 (26 comparable)	2.33 → 1.52	3.44 → 3.48	10.23 → 8.97	13.60 → 11.51
150 random library presets (44 comparable)	1.19 → 0.98	2.18 → 1.89	3.64 → 3.64	4.75 → 4.75
120 presets that draw motion vectors (30 comparable)	1.64 → 1.64	2.54 → 2.49	3.86 → 3.86	5.25 → 5.00
#750's 21 presets, 1440×896				4.25 → 4.25

• Wrap set, all 150. The frame-59 median fell 21.1 → 15.5; 41 presets are better by more than 2 and 9 worse.
• Single wins:
  • "liquid firesticks" f2: 26.7 → 2.3
  • "Tricolors" f30: 71.0 → 8.8
  • "Furry Synthesis" f59: 40.7 → 24.2
  • "Accelerator" f59: 26.4 → 14.7
  • "truth, rarity edit05" f30: 23.8 → 8.4. It now has MilkDrop's colour and wire mesh.
  • "fed - DSC-011 LSD Edit 4" was black and now shows MilkDrop's coloured fields.

Looked at frame by frame:

• "out of the paradigm" f30: 67.3 → 72.7. Its brightness is closer to MilkDrop and there is no artifact. It samples noise; see below.
• "1200773722134" f59: 4.2 → 8.5, a few percent darker. The blur chain alone causes it, and the chain now reads what MilkDrop reads.

Tried and dropped: full light for motion-vector segments. All 21 of #750's presets came out equal or worse.

The feedback rounds to 8 bits, as MilkDrop's does

With the blur fix alone, "Dead CellWerx" went 17.3 → 20.1 at f59, with a stepped ring and a red background. The root is buffer precision. MilkDrop's buffer is 8-bit and rounds every write to the nearest 1/255; butterchurn and projectM are 8-bit too. Our buffer is 10-bit (RGB10_A2).

Measured in patched MilkDrop with synthetic warps:

warp	MilkDrop	10-bit	now
+0.0025/frame, f59	59	39	60
+0.0015/frame, f59	0	19	0
×0.9 decay floor, warp shader	4	1	4
×0.9 decay floor, no warp shader	4	1	4

• What changed. In accurate mode the warp's output (preset shader and shader-less) is rounded to 1/255.
• What didn't. The buffer stays 10-bit, so waves, shapes and lines drawn after the warp keep their soft edges on screen. The comp pass isn't fed back and isn't rounded.
• "Dead CellWerx" f59: 20.0 → 8.7, with no ring and no red background.
• A plain 8-bit buffer was measured and rejected. It rounds those soft edges away in the feedback and darkens line presets:
  • "EoS - 7th galaxy" 43.8 → 63.3;
  • the motion-vector median at f30 3.86 → 4.72.
• Threshold. Round-to-nearest is used. On this card MilkDrop's threshold sits between 0.52 and 0.6, but D3D lets cards differ.

Before → after the rounding, same MilkDrop frames:

set	f10	f59
shader + bTexWrap = 0 (26)	3.48 → 2.84	11.51 → 11.11
random 150 (44)	1.89 → 1.86	4.75 → 4.68
100 without a warp shader (68)	1.24 → 1.24	mean −0.06, none worse by more than 0.67
motion vectors (30)	2.49 → 2.54	5.00 → 4.73
#750's 21		4.25 → 4.19

Main → branch, all three fixes, noise-free comparable presets:

• wrap set f30: 7.13 → 3.62;
• random 150 f59: 3.93 → 3.60;
• motion vectors f10: 2.16 → 2.00.

Nothing is worse by more than 1.03 except "fed - DSC-011", which went from black to MilkDrop's coloured fields.

Noise presets. BeatDrop seeds its noise textures from std::random_device and the performance counter (plugin.cpp:645-646, 2863). Two MilkDrop runs of a noise preset differ by up to 17.6, while noise-free presets come out byte-identical. Every main → branch delta above uses one copy of MilkDrop's frames for both sides.

Checks

• Tests: npm test passes 3020 of 3020, with 10 new tests.
• Isolated smoke (--user-data-dir): PASS. The real presets are unchanged; settings.json changed during the run because the installed app was open and writing it, while the smoke used its own profile.

Refs #752, #560, #695

[cayadraft]

@CaYatur

CaYaDraft — pull request

Summary
MilkDrop: fresh blur levels and MilkDrop's sampler wrap rules (#752)

Files
CHANGELOG.md (modified, +10/-0)
@@ -38,6 +38,16 @@ GitHub. The full reasoning and the measurements behind each change are in [ROADM

• Measured against patched MilkDrop on 120 presets that draw motion vectors, over the presets
  without user textures or randomness: the median frame-59 difference fell from 28.1 to 5.3, with
  no preset worse at frames 2, 10, 30 or 59.
  +- **Blur levels 2 and 3 are no longer one and two fr…

README.md (modified, +2/-2)
@@ -17,7 +17,7 @@ layers and 40 GPU effects, an OBS overlay, Spout / Syphon, **project


[![Platform](https://img.shields.io/badge/platform-Windows%20%…

README.tr.md (modified, +2/-2)
@@ -18,7 +18,7 @@ destesi**, ve MCP üzerinden yapay zekâ kontrolü.


[![Platform](https://img.shields.io/badge/platform-Windows%20%7C%20macOS%20%7C%20Linux-111997.sv…

ROADMAP.md (modified, +30/-2)
@@ -130,10 +130,10 @@ npm test
npm start -- --smoke

-- **3010 unit tests, all passing** on `main`. 703 of those shipped in v3.1.0;
+- **3015 unit tests, all passing** on `main`. 703 of those shipped in v3.1.0;
  105 came with v3.1.1; 163 came with v3.1.2; 157 came with v3.1.3 — 1128 at
  that tag — 469 more with v3.1.4, most of them from the MilkDrop work — 1597 at
-  that tag — 1402 mo…

src/visualizer/modes/milkdrop.js (modified, +46/-3)
@@ -364,10 +364,16 @@ uniform float uBias;
uniform vec3 uEdge;      // kenar karartma: (1-b1ed, b1ed, 5.0); kapaliyken (1,0,5)
uniform float uSat;      // 1: sonuc [0,1]'e doyuyor (MilkDrop'un 8 bit hedefi, #580)
void main(){
-  vec3 c = texture(uSrc, vUV).rgb * uCenter;
+  /* KAYNAK HER ZAMAN TEMEL SEVIYEDEN (textureLod 0). Bir onceki kademenin
+     \`out\` dokusu mipmapli ve mipmap kademele…

tests/milkdrop-blur-edge.test.js (modified, +10/-0)
@@ -56,6 +56,16 @@ test('shader: karartma karekök eğrisiyle uygulanıyor', () => {
  assert.match(FRAG[1], /outColor = vec4\(uSat > 0\.5 \? clamp\(r, 0\.0, 1\.0\) : r, 1\.0\);/);
});

+test('shader: kaynak her zaman temel seviyeden okunuyor (bayat mipmap yok)',…

**Note**
The local model is unavailable, so this is only the title and file list.

---
Stopping here. I will not merge, push, or change this pull request until you ask.

`/caya explain` describes the diff. `/caya plan` previews the next edit. `/caya revise` adds a commit.

<!-- cayadraft -->