Replace Dye Mute and Lab Vibrance with a signed Dye Separation

[marcinz606]

Supersedes #675 (@thetalkingdrum's draft), and builds directly on its four prototype commits, which are included in this branch. Extends #667.

What changed

#675 proposed two controls: Density Vibrance (a signed per-pixel spread rescale) and Spatial Damping (a toggle applying Dye Mute's strength through the same per-pixel mask). Reading them side by side, they compute the same thing.

Spatial damping:

d_mask = 2*sigmoid(spread/0.4) - 1
d_k    = 1 + (damp_global - 1) * d_mask,   damp_global = max(0, 1 - D)
       = 1 - min(1, D) * d_mask

Negative vibrance:

s = 2*sigmoid(spread/0.4) - 1
k = 1 + V*s = 1 - |V|*s

Same scale constant, same mean-preserving rescale of e_i = dens_i - d_min_i. Setting V = -D gives bit-identical output. So the toggle was a second spelling of the slider's negative half.

This collapses them into one control, Dye Separation (exposure.dye_separation, -0.5..0.5, 0 = off).

The sign selects which pixels the mask hits rather than just the direction: positive uses 1 - s and acts on pixels whose three dye densities sit close together, negative uses s and acts on already-separated ones. Neither end touches the other's population, which is what makes it read as selective rather than as a saturation slider. Flipping only the formula's sign would push and pull the same muted pixels and never reach the separated ones.

Removed

• density_saturation_damping and grade_saturation_damping() — Dye Mute's frame-uniform, grade-coupled damp, on both the CPU and GPU paths
• density_damping_spatial — the toggle that only chose between uniform and per-pixel
• Lab Vibrance — apply_vibrance, the model field, the shader branch, the slider, and its Alt+4 shortcut pair. Its job was selective saturation; Dye Separation now does that in density space, where it responds to the paper profile.
• Draft: density-space Vibrance/anti-vibrance + spatially-modulated Dye Mute (extends #667) #675's A/B comparison row, which existed to compare against Lab Vibrance

Net: +279 / -369 across 35 files.

The risky part

Lab Vibrance had a field in LabUniforms, declared identically in six shaders (lab, lab_sharpen_h/v, rl_blur_h, rl_div_v, rl_mult_v). Removing it shifts the offset of every field after it, so the struct.pack in gpu_engine.py drops to nine floats and a _pad3 holds the block at 48 bytes. A mistake here would silently corrupt sharpening and denoise parameters rather than fail loudly.

tests/test_pipeline_parity.py and tests/test_gpu_curve_parity.py both ran against a real GPU (77 and 7 passed, no skips), so CPU and GPU agree on the new layout.

CLAUDE.md claimed LabUniforms lives in 7 shaders including rl_init.wgsl. It binds no uniform at all — the auto layout prunes it — so that is corrected to 6.

Compatibility

Retired keys drop silently through migrations.py, so no saved edit warns on load. density_vibrance renames to dye_separation rather than dropping, so edits saved during the prototype keep their look. Default is 0.0 (identity), so no golden values shifted.

UI

Dye Separation sits in the Tone panel where Dye Mute was, beside Print Saturation. Chroma Denoise moves up into the Lab panel's COLOUR row to fill the slot Vibrance left; CLAHE is now full-width in DETAIL. Alt+4 / Alt+Shift+4 are unbound rather than renumbered, so existing muscle memory for the other Lab keys still works.

Docs

PIPELINE.md §3 and §6 and USER_GUIDE.md §6.2 and §7.1 updated. An audit of every documented slider range and constant against the code also turned up four errors that predate this branch:

• Fine Rotation is ±45°, documented as ±5° (FINE_ROTATION_LIMIT = 45.0)
• Print Saturation is 0.5–1.5, documented as 0.3–1.6
• Export offers neither ACES nor XYZ — EXPORT_COLOR_SPACES excludes both, but the guide listed them
• The Sensor Calibration panel was entirely undocumented, and two override.toml keys were missing

All fixed here.

Testing

• tests/test_dye_separation.py — identity no-op, both directions change output, the two signs are not mirror images, B&W inertness, and that the retired keys load without warning
• tests/test_gpu_curve_parity.py — CPU/GPU parity for both signs
• tests/test_saturation_damping.py deleted; its one surviving guard (the Lab stage ignoring print_slopes) moved to tests/test_density_saturation.py

make lint, make type, and the full suite are green: 2728 passed, 5 skipped.

Open question

dye_separation_spread_scale = 0.4 is still #675's unvalidated starting guess for where "muted" ends. It now drives one control instead of three, so tuning it no longer retunes unrelated behaviour, but it has not been checked by eye against real frames.