[kocean9-freedom]

After metric.corpus_score(..., n_bootstrap=10), reusing the same BLEU, chrF or TER object for ordinary corpus or sentence scoring leaves bs:10|seed:12345 in get_signature(), even though the new score has no confidence interval. A subsequent paired AR test without confidence estimation also inherits bs:10. The reported reproducibility settings therefore depend on earlier calls.

Reset the single-score bootstrap metadata after computing a new score, and clear inherited bs metadata before PairedTest adds its own settings. Scoring, resampling and p-value calculations are unchanged. Previously returned signature objects retain their original settings.

The regression tests compare reused and fresh metric instances across all three metrics, explicit/cached references, sentence scores, AR with and without confidence intervals, and paired BS. They also check that a later bootstrap call records its new count and seed. On unchanged master c596d9d, 12 new cases fail and six controls pass; all pass with the fix.

Validation on macOS with Python 3.12.14:

• 195 tests pass with python -m pytest --ignore=test/test_api.py --ignore=test/test_dataset.py; three additional offline API tests pass. This includes the existing WMT17 significance comparisons.
• mypy sacrebleu scripts test passes for 44 files. Ruff passes on the four changed files.
• Offline source-distribution and wheel builds pass. Their changed source/test files match the tested checkout.
• 33 public API scenarios pass from the unpacked wheel, including JSON score output and paired tests with one and two workers.

Dataset-download tests, the network-heavy shell suite and other platforms were not run. This does not include the independent fixes in #311 or #312.

AI assistance: OpenAI Codex performed the investigation, implementation, tests, self-review and description. No independent human review has occurred.