fix: clarify run detail result metrics

[pppolf]

Summary

• distinguish unscored answers, ordinary wrong answers, and execution errors in Run Detail
• expose Run-wide known input/output token subtotals and independent coverage on the paginated Responses API
• preserve protocol-v1 exact-token null semantics while showing explicitly incomplete evidence

Validation

• make lint
• make test (backend 951 passed, 33 skipped; frontend 47 passed)
• make smoke
• frontend production build
• docker compose config --quiet
• target Run verified through the local API and browser UI

No real Provider was called and no historical data or schema was modified.