A control's range is a claim about its input

valueRange is not a styling choice. It is an assertion that every value the control will ever be handed lies inside it — and the control is handed values by import, restore, migration and the previous version of your own code, not only by the user's finger.

/**
 * Deeper than the ±12 dB the bands span, because an imported profile's own trim is computed from
 * the summed response rather than from its tallest band, and so goes past −12: the lowest across a
 * sample of sixty published profiles was −12.1 dB.
 */
private const val PREAMP_MIN_DB = -15f

Slider(value = draft ?: stored, valueRange = PREAMP_MIN_DB..0f, …)

Traps

Out of range does not clamp the value — it clamps the thumb, and your readout keeps telling the truth. The widget positions the thumb from a fraction it computes and then coerces:

calcFraction(a, b, pos) = if (b - a == 0f) 0f else (pos - a) / (b - a)   … .coerceIn(0f, 1f)

So −12.1 on a −12f..0f range draws exactly like −12.0, while a readout rendering the stored value prints -12… or -12.1, depending on your rounding. The user sees a control sitting at its limit next to a number that is not the limit, and no exception is thrown anywhere.

Pressing the thumb where you can see it rewrites the value, and that is data loss with no prompt. The setter does not coerce, and nothing is emitted merely because the control became interactive — the emit is gesture-driven. A press sets the raw offset from the pointer position, and the handler is invoked with the value that position scales to, not with a clamped copy of the stored one. Since the thumb is drawn at the clamped position, a user pressing it exactly where the widget put it hands you the range endpoint, and your handler stores that over the out-of-range value. That is worse than a coercing setter, not better: the pinned thumb is what aims the gesture at the wrong number, so the widget looks correct right up to the moment it destroys the value it was showing. The user opened the screen to look at the setting and left having changed it. Nothing logs this; it is indistinguishable from a deliberate edit.

Sizing from the neighbouring control is the specific mistake. The bands span ±12, so ±12 looks like the obvious bound for the trim beside them. They are not the same quantity: one band's gain is one filter's gain, while a trim that makes room for all of them is computed from the summed response and legitimately goes deeper than any single band. Two controls in the same block, drawn the same way, measuring different things.

Measure the distribution before you pick the bound, and write the measurement into the constant. "The lowest across a sample of sixty published profiles was −12.1 dB" is worth ten times a round number, because it tells the next maintainer what would have to change for the bound to be wrong — and it is falsifiable. A bound with no recorded basis gets widened by guesswork on the next report.

Asymmetric is fine, and often correct. A trim whose purpose is to make headroom should not be allowed to boost; that would only move the clipping somewhere else. Do not "balance" a range for symmetry's sake — the range should follow what the value means, then be widened to cover what arrives.

If you genuinely must narrow the range, clamp at ingest and say so. Let the import reject or clamp, with a message, rather than letting the widget do it on first touch. A clamp the user was told about is a decision; a clamp the widget performs is a silent edit.

Round the readout the same way the control steps. A readout rounding to whole units over a continuous range shows the same number for two genuinely different states, so a user nudging the control sees nothing happen and nudges again. Either give the control discrete steps or print enough digits to distinguish them.

The bound belongs to the stored range, not the control. Anything else that reads the value — a second backend, an exporter, a settings summary — has to accept the same span. See one-setting-two-backends: a curve produced under one build's range and re-opened under a narrower one is the same failure arriving from your own past release.

Verifying it

Compute the extremes of the real input distribution and compare them to the declared range, rather than reasoning about it:

# every declared range in the UI
grep -rn --include='*.kt' 'valueRange' .

# the extremes of what actually arrives — here, the trim line of a sample of profile files.
# LC_ALL=C is load-bearing: under a non-C LC_NUMERIC, `sort -g` orders by the integer part only
# and reports a wrong extreme with no error (-12.1 before -12.9), which is the one thing this
# command exists to get right.
grep -h -i '^Preamp:' sample-profiles/*.txt \
  | sed -E 's/^[Pp]reamp:[[:space:]]*(-?[0-9.]+).*/\1/' \
  | LC_ALL=C sort -g | sed -n '1p;$p'

Then the two states that only appear with a value from outside:

• seed the store with a value one step outside the range, open the screen, and assert the rendered number equals the stored one — this is the check that catches the pinned thumb;
• do the same, then press the thumb and assert the stored value did not become exactly the range endpoint. Assert that, not "unchanged": a press anywhere emits the position it landed on, on a correct range too, so an unchanged-assertion fails on code that is right. Seeded at −12.1 against −12f..0f the press yields exactly −12.0; against −15f..0f the thumb sits at the true position and the emitted value stays within a pixel of −12.1. A test that only drags passes on a broken range either way, because dragging is supposed to change the value.