Expose the Contrast Mask's spacer as Mask Blur

[marcinz606]

The Contrast Mask's blur was pinned at 4% of the analysis grid. This exposes it as Mask Blur, beside Contrast Mask in the Tone panel.

Why it earns a control now

When the radius was cut from the first version, the only evidence I had was smooth synthetic content, where sigma 40 against 150 moved micro-contrast by 0.05. That did not justify a slider. Measured per octave on a real scan it clearly does, because the blur is not a strength but a frequency cut-off: it sets the scale above which tones are masked.

Band energy against unmasked, at gamma 0.5:

blur	2px	4px	8px	16px	32px	64px
1%	0.92	0.89	0.75	0.60	0.52	0.51
2%	0.93	0.93	0.89	0.77	0.61	0.54
4% (default)	0.94	0.94	0.94	0.91	0.79	0.64
8%	0.97	0.96	0.96	0.96	0.93	0.82

Narrower compresses harder (global range 0.50× at 1% against 0.62× at 4%, on a theoretical floor of 0.50) and lifts shadows sitting next to something bright, which is the mask line a printer sees on the sheet. At the narrow limit blur(val) → val and the operator collapses into a plain (1−g) reduction, which is Grade's job. 4% stays the default as the conservative end.

Note the first metric I reached for was blind to this: an RMS Laplacian only sees 1–2px, and even 1% blur is ~10px on a 1000px render, so micro-contrast barely moved (0.94 to 0.98) across the whole range. The per-octave decomposition is what showed the axis.

Naming

Mask Blur rather than Spacer, Unsharpness or Mask Sharpness.

• Spacer is the better darkroom name (it is literally the sheet holding mask off negative) but invites a thickness in mm, and mm needs FilmFormat from user-editable EXIF, already rejected as a fake unit on this feature.
• Mask Sharpness measures 100px against a 96px budget at a two-per-row slider's share, so it would push the label on hover, the defect just fixed on Tilt/Swing. It also inverts against Lab's Sharpening, which moves apparent detail the opposite way, and it would force the stored value from a measurable per-cent into an abstract knob.
• Mask Radius collides with Lab → Sharpening → Radius, also a blur radius.

Mask Blur is 62px, keeps the number a real per-cent, and needs no inversion.

Cache correctness

The gamma slider is a pure scalar on a cached plane, so the blur is the one part that genuinely rebuilds it. It now keys the plane cache in both engines and the GPU's mask texture. Verified by moving the blur under an unchanged source hash: CPU and GPU both re-render (8%→4% mean |Δ| 0.00242, 4%→1% 0.00618) and agree exactly, and parity at a non-default blur is 0.00000 max abs diff.

Dead without a mask, so it greys out at gamma 0, and it is global-only like the mask itself. It is deliberately not in _global_only, since that tuple's contract is "enabled exactly when global" and this has a second gate.

Verification

make all green, 3964 passed, 3 new tests (the cut-off behaviour, the cache rebuild under one source hash, and that it is inert with no mask). Driven headlessly through the real GUI on a real scan alongside the existing Contrast Mask checks. Both docs updated.