Update eval harness with richer scenarios and diagnostics

[zac]

Summary

• Expanded the built-in eval set with additional filesystem and catalog scenarios covering API shapes, recovery, and capability minimization.
• Added multi-step execute support plus clearer runtime/tool guidance so models use top-level returns and avoid unreturned async IIFEs.
• Tightened filesystem catalog metadata and added diagnostics/tests for missing return values and async IIFE misuse.
• Re-ran and updated both core and failures r5 baselines after clean live runs.

Testing

• swift test
• swift run --package-path Tools/CodeModeEval codemode-eval run
• Live codemode-eval llm runs for core --repeat 5 and failures --repeat 5
• Baseline comparison and report generation passed for both suites