Function pressure 

pub fn pressure(
    dist: &Distribution,
    side: Side,
    values: &[ParamValue],
    top_set_len: usize,
) -> Result<f64>

The crowding fraction on side: the share of the top set that band counts as “at the bound” there — §8.1’s pressure(side S) = |{ t in top set : t is in band(dist, S) }| / |top set|.

values and top_set_len count two different things, on purpose

values is one parameter’s recorded value from each top-set trial that recorded it — never the trials themselves, and never padded for the ones that did not record it (a trial that never took a conditional branch simply contributes no entry). top_set_len is |top set|: the size of the whole top set top_set produced, regardless of how many of its trials happen to carry this particular parameter. §8’s formula divides by the second number, never the first; the two coincide only when every top-set trial recorded the parameter, which is not true for one reached through a conditional branch.

This is a real, deliberate consequence for a conditional (define-by-run) parameter, not an edge case filed off: one that only 3 of a 16-member top set ever reached can score at most 3 / 16 = 0.1875 — never 3 / 3 = 1.0 — however tightly those 3 sit against the bound. A rarely-taken branch therefore cannot clear the pressure >= 0.5 firing threshold (§8) on the strength of the few trials that ever reached it, because top_set’s warmup gate (completed >= max(16, 4 * dim_k), §8.4) is global, not per-parameter: there is no separate, smaller warmup for a branch most trials never take. atune documents conditional parameters as a supported pattern, so a reader of this function will meet this case — it is the formula’s own reading of §8, not a gap in it.

Out-of-range values are the strongest evidence, not an error (§8.1)

A value can legitimately sit outside dist’s current range: a warm-started study’s seed trials may be drawn from a wider range than the study’s own declared bounds, and a veto rollback can narrow a bound under trials already recorded. Such a value counts as maximally in the band on the side it overshoots. It is never filtered out and never raises Error::OutOfRange — filtering would silently hide exactly the crowding evidence this rule exists to detect.

Returns 0.0, never NaN, when values is empty but top_set_len is not: the case above, where no top-set trial recorded the parameter. An empty top set itself (top_set_len == 0) is refused instead — see Errors.

Errors

• Everything band returns for dist.
• Error::InvalidSpace if top_set_len is 0 — there is no top set to measure crowding against — or if values.len() > top_set_len, since values must be a subset of the top set it was drawn from.
• Error::OutOfRange if a member of values is not the kind of value dist produces.

use atune_core::space::{Distribution, ParamValue, Side, pressure};

let lr = Distribution::float(0.0, 1.0).unwrap();
let top = [ParamValue::F64(0.99), ParamValue::F64(0.5)];
assert_eq!(pressure(&lr, Side::High, &top, top.len()).unwrap(), 0.5);
assert_eq!(pressure(&lr, Side::Low, &top, top.len()).unwrap(), 0.0);

// A value beyond the range still counts, on the side it overshoots.
let drifted = [ParamValue::F64(5.0)];
assert_eq!(pressure(&lr, Side::High, &drifted, 1).unwrap(), 1.0);

// A parameter only 1 of a 4-trial top set ever recorded scores at most
// 0.25, even sitting exactly on the bound -- see the note above.
assert_eq!(
    pressure(&lr, Side::High, &[ParamValue::F64(1.0)], 4).unwrap(),
    0.25
);