[BUG] dyck_language scores with inspect_ai's exact(), which ignored word order in every release up to 0.3.252

[LippInc]

Describe the bug

src/lighteval/tasks/tasks/dyck_language.py scores with inspect_ai's exact():

from inspect_ai.scorer import exact
...
scorer=exact(),

In every inspect-ai release from 0.3.27 (Sep 2024) through 0.3.252, exact() compared sets of words: order and token multiplicity were discarded before comparison (upstream UKGovernmentBEIS/inspect_ai#4618, fixed by #4619, first released in 0.3.253). Dyck targets are space-separated sequences of closing brackets, and order is the property being tested, so under inspect-ai <= 0.3.252 a wrongly-ordered closing sequence with the right bracket types scores 1.0.

To Reproduce

On inspect-ai 0.3.252:

from inspect_ai.scorer._classification import max_exact_score

print(max_exact_score("} ] )", [") ] }"]))              # wrong bracket order
print(max_exact_score("hello world", ["world hello"]))  # word order
print(max_exact_score("hello world", ["hello world"]))  # control: identical
print(max_exact_score("bananas", ["hello world"]))      # control: unrelated

1.0
1.0
1.0
0.0

Same script on the 0.3.253 wheel: the first two print 0.0, controls unchanged.

Expected behavior

A wrongly-ordered closing sequence scores 0. Requiring inspect-ai >= 0.3.253 for these tasks (or otherwise ensuring the fixed scorer is used) would do it. Happy to open a PR for the pin bump. Results produced with older pins can only be inflated, never deflated, because the old comparison marks more answers correct.

Version info

lighteval main; inspect-ai 0.3.252 vs 0.3.253; Python 3.13.5, Windows 11.

(Claude helped write this up; I ran the repro myself, on both wheels.)