A world-space param is not bounded by its clamp — the frame is the bound

The section above is about a band term that stops moving. This one is about a band term that moves too far, and it is the opposite failure with the same cause: a gain written against the wrong scale.

clamp(bass * 0.45, 0, 0.3) cannot leave the frame. It is a unitless multiplier into brightness, or width, or a hue offset, and the 0.3 is a real ceiling. But some params are world-space: they multiply a band into a coordinate. scale on a spectrum readout is a world height per unit of band level; span, baseline, size on the attractor, and any bin()- or index-driven geometry are the same kind of quantity. Write

scale = "3.20"     # spectrum, [spectrum] layout = "polyline"

and a fully-driven element stands 3.2 world units tall. The visible half-height is 1.0. The figure is not clipped, not squashed and not dimmed — it is outside the picture, and there is no clamp anywhere in that line for a reviewer, a reachability walk or --report’s occupancy column to catch it on.

This shipped. spectrum_ridge carried scale = 3.20 from before ADR-0049 normalized the bands to 0..1. Afterwards the same constant multiplied a value roughly five times larger, and the preset rendered as an empty frame under --signal noise:7 for its entire life while every automated gate stayed green. The repair was 3.20 -> 0.60, chosen so a fully-driven element lands just inside the frame.

So when a band drives a coordinate, do the arithmetic in world units before you render anything:

at the loudest the band can go, where does the element land? If the answer is past 1.0 vertically or past your span horizontally, the preset has an invisible top end however good it looks at rehearsal level.

Two things that will not save you, both learned the hard way:

• A clamp on the term does not make it safe unless its upper bound is itself in world units and inside the frame. clamp(bass * 3.2, 0, 3.0) is still off frame; the clamp bounds the number, not the picture.
• Looking at it at a normal listening level does not save you either. The defect lives at the top of the range. The retired spectrum_comb and spectrum_corona clipped their tallest bars off the top edge on every beat and still looked fine, because a comb roots each bar on a shared baseline and only the tips leave.

Check it at the top of the range, not in the middle:

cargo run -p standalone --example shot -- --preset-file presets/yours.toml --signal noise:7

core/tests/sanity.rs catches the total case — a figure so far out that the frame comes back empty. It suppresses the backdrop, so a bg_vignette cannot stand in for a figure that is not there, and it measures against the frame’s own ground rather than against black — the mean tone of its most populous luminance band (ADR-0126), so a scene that paints its own paper is measured on what it drew instead of reading as a completely full frame. It also prints a per-preset excitation ratio (coverage when fully driven over coverage at a moderate level) on every run. That ratio is a report rather than a gate, and the reason is worth knowing: a partial over-scale that clips only the tips costs almost no pixels, so the number comes back healthy. Nothing automated will catch a tip that leaves the frame. That one is yours to check.