fix(eval): separate parity from score eligibility

[vincentkoc]

What does this PR do?

Reports scoped Harbor parity independently from native leaderboard eligibility, so complete native runs remain scoreable without an unsupported parity claim.

Why?

The scoped-parity change in #46 accidentally made parity validation a hard eligibility gate. Reprocessing the repaired 93-run matrix then validated every archive but selected 0/93 runs, all with parity_not_validated. The ShellBench matrix contract gates eligibility on coverage, infrastructure health, harness-wide failures, model identity, and trace completeness; parity is separate audit metadata.

Changes

• remove parity validation from native score eligibility
• retain per-run parity status and scope
• add pair-level parity counts and unvalidated run labels
• show parity coverage separately in the cleaned leaderboard
• update parity regression expectations

Tests

• PYTHONPATH=. .venv/bin/pytest -q tests/test_native_eval*.py (99 passed)

• .venv/bin/ruff check scripts/native_eval tests/test_native_eval*.py

• repaired seed replay: 93 selected, 93 eligible, 0 parity-validated, no exclusion reasons

• python -m pytest -q equivalent affected native suite passes locally

• python -m ruff check clawbench app.py scripts tests affected native surfaces pass locally