Bounding shrink work

By default shrinking runs until no smaller candidate still fails, re-running the property once per accepted step. On expensive properties or very large inputs you can cap the number of accepted shrink steps with maxShrinks:

#[Property(runs: 200, maxShrinks: 25)]

maxShrinks: null (the default) means no cap. maxShrinks: 0 disables shrinking entirely and reports the original counterexample unchanged. The cap counts accepted shrink steps, not test executions.

Bounding the work, not the result

That last sentence is the limitation: a descent spends most of its time on candidates it rejects, and maxShrinks never counts those. On a large collection the rejected candidates alone can cost more than the random phase that found the failure. The engine's PropertyConfig bounds the descent from the other side:

use Rasuvaeff\PropertyTesting\Runner\PropertyConfig;
use Rasuvaeff\PropertyTesting\Runner\ShrinkMode;

new PropertyConfig(shrink: ShrinkMode::Off);   // report the counterexample as generated
new PropertyConfig(shrinkBudgetMs: 500);       // descend for at most 500 ms

ShrinkMode::Off skips the descent: no trial, no shrink event, zero steps and zero trials on the counterexample. A shrinkBudgetMs resolves to ShrinkMode::Bounded, which stops the descent on wall clock and reports the best candidate it reached — not the original one.

The budget is a millisecond count of at least 1, and at most large enough to still convert into the nanosecond deadline the runner compares the clock against. Past that point the arithmetic leaves the integer range and the deadline stops being one, so such a budget is rejected at construction rather than silently disabled.

What a budget costs

A wall-clock budget is the one knob here that gives up determinism. How far the descent gets depends on how long the property body takes, so the same seed can minimise to a different counterexample on a fast machine and a slow one. That is a deliberate trade: the budget answers "the descent hung for a minute", not "reproduce this exactly". For reproduction, pin the seed or let the regression corpus replay the recorded input — both stay deterministic.