Here's an idea that closes #2180 (cc @aarchiba): consider a test which asserts that the largest integer is 100. Obviously this will fail, but shrinking will make it look like an off-by-one error!

@given(st.integers(min_value=0))
def test_threshold_problem(x):
    target(float(x))
    # target(float(x * 2), label="another target")
    # assert x <= 100000
    assert x <= 100

Fortunately, reporting the highest observation seen by target makes it easy to tell that while x=101 was the minimal counterexample, much larger errors exist:

Highest target score: 9.14676e+08  (label='')

Falsifying example: test_threshold_problem(
    x=101,
)
Traceback (most recent call last): ...

And if we un-comment our other target calls and assertions, we can see that this works smoothly with multi-objective optimisation and reporting of multiple errors too:

Highest target scores:
     3.27686e+06  (label='')
     6.55371e+06  (label='another target')

Falsifying example: test_threshold_problem(
    x=100001,
)
Traceback (most recent call last): ...

Falsifying example: test_threshold_problem(
    x=101,
)
Traceback (most recent call last): ...

As in #2234, either there's a threshold (in which case the maximum is always from a failing test) or there isn't (and the score is irrelevant). For consistency we therefore report the maximum score seen from any example and trust that users can skip over those which are not relevant. While more verbose, I expect this to be a substantial improvement for scientific and data-oriented users where target()ing an error budget is a very common style of test.