Add a single benchmark artifact verification path

[yinli-systems]

What changed

• add l20-codeforge verify-artifacts with structured PASS/FAIL results, missing/mismatch reporting, and repository-root containment
• verify all five claim-bearing artifacts documented in REPRODUCIBILITY.md
• replace the shorter inline CI hash script with the same package command users run locally
• add regression tests for valid, changed, missing, and escaping artifact paths
• update canonical repository links and make README/test expectations resistant to normal test-count growth

Why

The reproducibility guide listed five hashes, but CI checked only three using a duplicated inline map. That allowed evidence drift between the documented and automated verification paths.

Impact

README claims, local verification, and CI now share one manifest and one implementation. The benchmark numbers and artifacts are unchanged.

Checks

• PATH="$PWD/.venv/bin:$PATH" python -m pytest -q — 137 passed
• python -m l20_codeforge verify-artifacts — PASS, 5/5 hashes matched
• python scripts/build_generalization_scorecard.py --output-dir <temp> — PASS, all 9 checks passed
• ruff check src/l20_codeforge/artifacts.py src/l20_codeforge/cli.py tests/test_artifacts.py — passed
• git diff --check — passed

Full hidden LiveCodeBench replay and GPU/model execution remain outside this documentation/verification change.