Boundary bias

Uniform random sampling wastes most of its draws on the uninteresting middle of a range. Bugs cluster at the edges — 0, ±1, the exact bounds — so the numeric generators deliberately skew toward them instead of sampling uniformly.

How it works

Gen::int(), Gen::intBetween(), Gen::float() and Gen::floatBetween() roll a 1-in-5 chance on every draw (IntArbitrary/FloatArbitrary's BIAS_DENOMINATOR = 5). On that roll, instead of a uniform value they return one of the in-range boundary candidates computed by the internal Boundary helper:

• Integers: the distinct values among 0, 1, -1, $min, $max that fall inside [$min, $max]. $min and $max are always in range by construction, so the candidate list is never empty — a range like intBetween(10, 20) still biases toward its own 10 and 20 even though 0/1/-1 are all out of range.
• Floats: the distinct values among 0.0, $min that fall inside [$min, $max). FloatArbitrary generates a half-open range, so $max itself is deliberately excluded from both the boundary list and ordinary generation — it can never be emitted as a value. A degenerate range where neither 0.0 nor $min qualifies (empty boundary list) just falls back to uniform sampling for that draw.

The other 4 draws in 5 sample uniformly across the full range, so the generator still explores the interior — bias shifts the odds, it doesn't replace coverage.

Shrinking is unaffected

Boundary bias only changes what generation returns; it says nothing about how a value shrinks. Both arbitraries shrink toward the same target regardless of how a value was produced — IntArbitrary by halving the distance to max($min, min($max, 0)), FloatArbitrary with the single candidate max($min, min($max, 0.0)). A boundary-biased draw and a uniformly-sampled draw with the same value produce identical shrink trees.

Why this matters for property design

• Don't write Gen::filter($ints, fn($n) => $n !== 0) expecting the filter to rarely trigger — with a 1-in-5 boundary roll landing on 0 whenever it's in range, plus 0 being a frequent target of ordinary shrinking, that filter discards far more often than a uniform-only model would suggest. Prefer Gen::intBetween(1, $max) (construct the domain) over filtering zero out of it.
• A property that "usually passes but sometimes fails at the edges" is exactly what this bias is built to surface fast — you don't need runs: 10000 to hit min/max/0; a few hundred runs already samples them repeatedly.
• Gen::floatSpecial() is the separate, opt-in mechanism for NAN/±INF/ -0.0 — the default float()/floatBetween() bias stays entirely in-range and finite on purpose. See the generators overview for the full catalog.