Fix perf gate crash when report uses nested metrics schema

[Navy10021]

Motivation

• The perf gate script crashed with a TypeError when the evaluation report used the newer JSON schema that nests metrics under metrics, because the script expected flat top-level fields and tried to format None values.
• The goal is to make scripts/check_perf_gate.py robust to both the v1 nested schema and legacy flat schema and to avoid formatting errors for missing metrics.

Description

• Change _extract_metrics to read from report.get("metrics", report) so the v1 nested metrics block is used when present while falling back to legacy flat keys.
• Add a small helper _fmt_metric(value, spec) that returns "N/A" for None and otherwise formats the value, and use it for verbose output to prevent NoneType formatting errors.
• Update verbose printing to use _fmt_metric for blue_win_rate and cvar_10, while preserving gate logic that treats missing metrics as failures with user-friendly messages.

Testing

• Ran python scripts/check_perf_gate.py /tmp/perf_gate_sample.json --min-win-rate 0.45 --max-cvar10 40.0 with a v1-style report and observed normal evaluation output and expected gate failure due to thresholds (script ran without crashing).
• Ran python scripts/check_perf_gate.py /tmp/perf_gate_empty.json --min-win-rate 0.45 --max-cvar10 40.0 with a report missing metrics and observed N/A printed and clean failure messages instead of an exception.
• Ran the end-to-end flow python evaluate.py --fast --no-progress --seed 42 --output-json /tmp/perf_gate.json followed by python scripts/check_perf_gate.py /tmp/perf_gate.json --min-win-rate 0.45 --max-cvar10 40.0 and confirmed the perf gate runs to completion (no TypeError) and reports gate result (failed on CVaR threshold)

---

Codex Task