skill-creator eval viewer: Benchmark tab empty due to configurations[] format not normalized

[mr-magaia]

Bug Report: Benchmark tab renders empty in eval viewer

Summary

The Benchmark tab in the skill-creator eval viewer renders empty when benchmark.json uses the configurations[] array format (which is the format Claude agents commonly produce). The viewer's viewer.html expects a different nested format with run_summary and runs[].

Root Cause

generate_review.py embeds benchmark.json as-is without normalizing the schema. viewer.html (line ~1122+) expects:

• data.run_summary — a dict keyed by configuration name, each value containing {mean, stddev} stat objects
• data.runs[] — array of objects with configuration string and nested result.pass_rate

But Claude agents producing benchmark.json commonly output:

{
  "configurations": [
    {
      "name": "with_skill (v2)",
      "pass_rate": {"mean": 1.0, "stddev": 0.0},
      "tokens": {"mean": 27973, "stddev": 4157},
      "time_seconds": {"mean": 56.5, "stddev": 12.2}
    }
  ],
  "delta_v2_vs_baseline": {
    "pass_rate": "+11.1% (100% vs 89%)",
    "tokens": "+68%"
  },
  "analyst_observations": ["v2 skill maintains 100% pass rate..."]
}

Fix

Add a normalize_benchmark() function in generate_review.py that converts between formats:

1. configurations[] array → run_summary dict (keyed by name)
2. Auto-generate runs[] from configurations if missing
3. config → configuration field rename in existing runs
4. Flat stat values → {mean, stddev} objects
5. analyst_observations accepted as alias for notes

Reproduction

1. Run skill-creator eval workflow
2. Aggregate produces benchmark.json with configurations[] format
3. Run generate_review.py --benchmark benchmark.json
4. Benchmark tab is empty — no data rendered

Related

This is the companion issue to #604 (Formal Grades display broken). Both are schema mismatch bugs between generate_review.py output and viewer.html expectations.

Environment

• skill-creator plugin (commit bd041495bd2a)
• Python 3.12, macOS