test(llm-coverage): fix and unskip the backup-judge test (#48)

[serkancakmakk]

Summary

Closes #48.

The "should use backup judge when one primary judge fails" test in src/cli/evaluators/__tests__/llm-coverage-evaluator.test.ts was .skipped because its expectations were stale — they were written for an older DEFAULT_JUDGES set (qwen / gpt-oss / glm) that no longer exists. This rewrites the test against the current DEFAULT_JUDGES and un-skips it.

What changed

Current DEFAULT_JUDGES (all holistic): google/gemini-2.5-flash, openai/gpt-4.1-mini, anthropic/claude-haiku-4.5; backup judge: anthropic/claude-haiku-4.5 (id backup-claude-4-5-haiku).

The rewritten test:

• Fails one primary judge (gpt-4.1-mini) and lets the other two succeed, so successfulJudgements < totalJudgesAttempted triggers the backup phase.
• Matches the backup judge by its id (backup-claude-4-5-haiku) first, since it shares its model with a primary judge and couldn't otherwise be distinguished in the mock.
• Averages the two surviving primaries plus the backup: (0.8 + 0.6 + 0.4) / 3 = 0.6, matching the source's avgScore.toFixed(2) consensus logic.

The assertions still verify the meaningful behavior — backup activation (DEFAULT_JUDGES.length + 1 judge calls), the consensus score, the "NOTE: Backup judge was used to supplement failed primary judges." reflection, and no error — so it reflects intended behavior rather than being forced to pass.

Acceptance criteria

• The "backup judge" test is un-skipped and passing
• Test reflects the actual intended backup-judge behavior
• llm-coverage-evaluator suite green

Verification

pnpm exec vitest run --project cli src/cli/evaluators/__tests__/llm-coverage-evaluator.test.ts
# Tests  39 passed (39)   ← 0 skipped

No .skip remains in the file.