fix(sampling): preserve cached choices during scoring

[Excelius-Wang]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

When sampling from a valid cache, generate returns the cached choices list. do_sample appends the ground truth to that same list, changing the cache's answer count. Reusing the cache key then misses the exact-count check and attempts generation; with --sampler_engine no, this raises AttributeError: 'NoneType' object has no attribute 'infer'. Duplicate rows within a batch also share the mutated list.

Copy the choices list before adding the ground truth. This preserves the cache while reusing the existing scoring and output logic.

Experiment results

CUDA_VISIBLE_DEVICES='' OMP_NUM_THREADS=1 HF_HUB_OFFLINE=1 \
  PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest \
  tests/general/test_sampler_cache.py \
  tests/general/test_sampler_engine_kwargs.py -q
pre-commit run --all-files
pre-commit run isort --all-files

• 19 tests pass, covering repeated calls, duplicate rows within a batch, cached and generated choices, ORM/PRM filtering, unchanged inputs/cache, and output contents.
• Before the fix, all four cached single-row regression cases fail on the second call; the four uncached controls pass.
• A separate CPU probe generated a cache through SwiftSampling.run, then reused it with the original dataset sampled twice via #2. The upstream method fails on the second row; the fixed method writes both scored rows and preserves the cache.
• All pre-commit hooks pass.

Sampling and cache methods run natively; model outputs and reward scores are mocked. Full model inference and a live Ray cluster were not run.