fix(benchmarks): apply limit after seeded shuffle

[Achordchan]

Summary

Benchmark runs now load the full filtered question set before applying the seeded shuffle and --limit. Different seeds can therefore select different question subsets, while --no-shuffle continues to select the canonical prefix.

Fixes #32

Changes

• stop truncating the dataset inside load_questions() before shuffling
• retain the existing post-shuffle limit as the single selection boundary
• add regression coverage for seed-dependent sampling and canonical no-shuffle ordering
• document the fix in the unreleased changelog

Verification

• .venv/bin/pytest tests/test_benchmark_runner_sampling.py -q — 2 passed
• .venv/bin/ruff check benchmarks/public/runner/run_subprocess.py tests/test_benchmark_runner_sampling.py — passed
• .venv/bin/python tools/import_smoke.py — 338/338 modules imported
• full .venv/bin/pytest -q — 1707 passed, 4 skipped, 2 failed outside this change
  • the network failure also reproduces on untouched origin/main because this environment resolves www.iana.org to reserved address 198.18.1.21
  • the TUI timing failure passed on isolated rerun

The repository-wide Ruff baseline currently reports existing formatting and lint findings; the modified Python files pass ruff check, and the new test file passes the formatter check.

AI assistance was used during implementation. I reviewed the final diff and ran the verification listed above.

[dq-ai-dev]

The regression coverage makes sense, but could we combine these into one parameterized test for the two seeds and --no-shuffle, sharing the setup? A loader stub that honors limit would let the selected-ID assertions catch the original bug directly, without also asserting that limit was omitted from the call.

[Achordchan]

Thanks, that makes sense. Updated in 2cff817: the two seeds and --no-shuffle now share one parameterized test, and the loader stub honors limit. The assertions check the selected IDs directly; the assertion about the loader call is gone.

All three cases pass, and Ruff passes. I also checked the regression by restoring the original limit-before-shuffle behavior in the test process: both seeded cases fail on the selected IDs, while --no-shuffle still passes.

[dq-ai-dev]

Thanks @Achordchan, this looks good to go. Merging...