test(evals): fix anomalies and expand eval coverage

[dzianisv]

Summary

• Fix 4 anomalies in existing eval assertions that were misleading or too loose
• Add 19 new eval test cases across all 3 eval configs (judge, stuck, compression)
• All evals pass: judge 31/31, stuck 18/18, compression 17/17
• Unit tests: 319 passed, 5 skipped

Anomalies Fixed

File	Issue
promptfooconfig.yaml #19	Description said "COMPLETE" but assertion expected complete === false
stuck-detection.yaml "Task finished"	Loose assertion accepted reason: "working" as passing
stuck-detection.yaml "Very short delay"	Tautological stuck === false || shouldNudge === false always passed
post-compression.yaml #2/#3	Accepted continue_task when needs_github_update was the correct answer

New Test Cases

Judge eval (8 new, 23→31 total):

• Agent stopped mid-task with pending TODOs (issue Session stopped #109 scenario)
• Agent claims fix but test output shows warnings
• Agent in retry/fix loop (3 attempts, still failing)
• Partial implementation (created files but didn't wire up)
• Gold-plating but task completed — COMPLETE
• Context window exhaustion (agent says "continuing" then stops)
• Agent forgot to run requested tests
• Agent commits directly to main violating workflow rules

Stuck detection (6 new, 12→18 total):

• Retry loop (same failed command repeated)
• Long-running build (150s) — NOT stuck, just slow
• Incomplete message with short delay
• Planning-only tokens, no action taken
• Rate-limited agent — should NOT be nudged
• Stuck agent must NOT be classified as complete

Post-compression (5 new, 12→17 total):

• PR with failing CI
• Agent debugging test failures mid-stream
• Multiple PRs open, agent working on one
• Task blocked on missing secrets
• Force-push scenario with rewritten history