Constant SATURATED_OCCUPANCY 

pub const SATURATED_OCCUPANCY: f32 = 0.9;

Occupancy at or above which a clamp() is reported as Saturated — the fraction of hops its inner value may spend pinned at the upper bound before the binding stops being a function of the audio and becomes a constant (ADR-0062).

A measured constant, not a principled one. Plan 0056 Phase 3 measured both sides of it — the library that passes, and the library that should not — over 339 clamped bindings each, on the 12 s dynamic:110 probe:

occupancy      today   pre-retune (80c5dff^)
[0.00, 0.10)      29        6
[0.10, 0.25)     171       11
[0.25, 0.50)     138       22
[0.50, 0.75)       1       51
[0.75, 0.90)       0      104
[0.90, 1.01)       0      145

The retuned library’s highest is 0.609 (Aurora.warp) and its next is 0.444, so 0.9 clears the measured maximum by 0.29 and the body of the distribution by twice that. The saturated library it exists to catch puts 145 bindings across 23 of 35 presets above it — the gate would have failed the build the day ADR-0049 landed.

Two things this value is not. It is not the most sensitive threshold that still separates the two libraries: 0.75 would catch 249 of the 339 pre-retune bindings rather than 145, but it would sit only 0.14 above a shipped, reviewed preset, and a HARD gate that fires on good content buys exemptions — which are the thing that dulls the instrument. And it does not see the marginal form of the defect: one binding pinned for 50-90 % of a track, in a preset with no severe case beside it, passes. What makes that acceptable is that the defect arrives in clusters — every affected preset in the pre-retune set carried a severe case too.

It has a shelf life. Re-measure it whenever the library changes materially, and expect to move it rather than to bless a preset through it.