MilkDrop fidelity: hue colour and the fixed composite as MilkDrop draws them

[CaYatur]

What

This fixes a regression from 52c7391 (#560) that #611 found, and checks the fixed composite's hue against MilkDrop's source: Nullsoft's D3D11 code (jecassis/foo_vis_milk2 5b44cea) and the D3D9 code (BeatDrop 53d83ee).

• Preset shaders get the full hue colour. MilkDrop gives every composite shader the full hue_shader colour, whatever the preset's fShader says ("since we don't know if shader uses it or not", milkdropfs.cpp:4122). fShader scales the colour only on the fixed pipeline and in the shader MilkDrop generates. 52c7391 applied the fixed pipeline's rule to shaders too, so with fidelity on:
  • 914 corpus presets that read hue_shader with fShader = 0 lost their colour;
  • 36 others got only part of it.
• Corners in the right place. MilkDrop weighs the four corners by the composite grid's screen position, with y = 1 on top (plugin.cpp:1475-1490). Our translation flipped y, which mirrored the colours top to bottom. A screen readback with four pure colours now puts corner 0 at the top right, as in MilkDrop; before this change it was at the bottom right.
• The fixed pipeline, drawn as MilkDrop draws it.
  • The amount is the file's fShader, blended linearly through a transition, and it is no longer clamped.
  • Each draw's corner colour (gamma × layer share × hue) goes through COLOR_NORM, so fShader = 10 wraps into other colours as in MilkDrop. fixedCompWeights replaces fixedGammaGain.
  • The quad is interpolated as MilkDrop's two triangles instead of bilinearly.
  • A generated composite stage no longer applies a partial fShader twice.
• Echo fade during a blend. The fade now uses the new side's blended echo alpha, as MilkDrop does, so the old echo fades out even when the new preset has none.

The legacy path keeps its own shader branch.

Measured

• Screen readback: shader corners sit where MilkDrop puts them. The fixed quad's centre is the mean of corners 1 and 2, as with MilkDrop's two triangles.

• Fixed pipeline render check: 72 of 72 cases within 2/255 of the model, including gamma 3.7 with echo 0.5 (eight draws). Rounding after every draw, as MilkDrop's 8-bit buffer does, differs from our single rounding by at most 0.6/255 in those cases.

• Against main (last 10 of 60 frames, up to 120 presets per group):

  Group	Result
  Shader presets that don't read hue_shader	120/120 identical
  fShader = 0 readers	95/120 changed; median saturation 0.006 → 0.150
  Full readers (orientation only)	90/120 changed, median 4.3%
  Partial readers	28/36 changed
  Fixed pipeline with fShader > 0.001	30/120 changed, median 0.5%; 12 of the 16 at 10
  Fixed pipeline without fShader	rounding only, largest 0.36%

• Fidelity off: the fixed pipeline and the control group are byte-identical. Presets that read hue_shader differ by one level in at most 0.1% of pixels, because four equal corner colours are summed in a different order.

• 900-preset sample: one preset moves from blown to clean. It scales its picture by hue_shader, which was white on main. Rendered alone, its brightness goes from 0.555 to 0.507.

Tests

• 3 new tests and 3 test files reworked; 18 of 18 mutations are caught.
• npm test: 2184 of 2184 pass, and each commit passes on its own. The changed files also pass under Linux emulation.
• npm run smoke: PASS, with the userData checksums unchanged.

Next

The remaining missing-key fallbacks, and MilkDrop's linear blending of its other file-only values. The orientation of uv in preset shaders is listed as an open check.

Refs #580
Refs #560