Function conditional_tail_percentile 

pub fn conditional_tail_percentile(
    art: &CalibrationArtifact,
    language: &str,
    metric: &str,
    value: f64,
) -> Option<f64>

Percentile of value within the corpus’s non-trivial tail for (language, metric) — the conditional lookup the anchored hotspot score uses in place of percentile.

The plain percentile compares against the whole pool, which for a complexity metric is dominated by trivial functions — the corpus median cognitive is 0 — so every real file saturates the top percentiles and the anchored score stops discriminating. This conditions the comparison on the functions that carry any decision structure (metric value above [TRIVIALITY_THRESHOLD]), re-spreading the informative upper tail across [0, 1]:

p0      = corpus fraction of the pool at value <= 0   (the trivial share)
cp_tail = clamp((cp - p0) / (1 - p0), 0, 1)            for value above 0
        = 0                                            for value <= 0

where cp is the plain percentile. p0 is read straight off the stored breakpoints by [trivial_share] as the highest index still holding a trivial value — the percentile the lookup itself assigns to the top of the zero plateau — so a value just above the plateau maps to cp_tail ≈ 0 and the corpus maximum to cp_tail = 1. It is pure arithmetic on breakpoints every artifact already carries: no format change, no corpus rebuild.

Returns None under the same conditions as percentile (unknown language / metric, or the language pooled below MIN_LANG_SAMPLE) and when the pool is entirely trivial (p0 == 1, an empty tail with no informative content) — an honest omission, never a fabricated value, matching the no-anchor path the caller already handles.