Stop choice extraction matching the last letter of a parenthesized word

[anugrah55]

sampling.extract_choice_response is what most model wrappers use in sample_choice to pull the chosen letter out of a reply. For replies longer than two characters it searches for \(?(\w)\), which also matches the last letter of any word that happens to be followed by ):

>>> from concordia.utils import sampling
>>> sampling.extract_choice_response('I think (probably) the answer is (b)')
'y'
>>> sampling.extract_choice_response('maybe) (a)')
'e'
>>> sampling.extract_choice_response('(yes) a')
's'

Usually the stray letter is not one of the offered choices, so the attempt is discarded and retried at a higher temperature. When the question has enough options that the stray letter is a valid key (for example e with five or more choices), the wrong answer is accepted without any error.

Fix

Require a word boundary before the choice character (\(?\b(\w)\)), so only a standalone character such as (b) or b) is accepted. The three replies above now give 'b', 'a' and None.

Callers are unchanged. The one- and two-character fast paths in extract_choice_response are not touched.

Tests

Three cases added to sampling_test.py. They fail on current main and pass with the change; the existing cases are unaffected.

Run on macOS arm64, Python 3.13.9, on top of 026de0c:

• pytest concordia: 1061 passed, 33 skipped
• pylint --errors-only and pyrefly check on the changed files: clean

[anugrah55]

Note on the red Test examples check: it fails in the "Typecheck examples" step on examples/concordia_island/run.py:647 (Tensor returned where np.ndarray is declared), a file this PR does not touch. The same step has failed on every push to main since 2026-10-06, so it is unrelated to this change. Test Concordia and the other checks pass.