feat: add multi-metric trend release gates

[brandonwise]

Summary

• extend argate trend beyond pass-rate drift to also report average latency, p95 latency, and average cost trends across historical runs
• keep existing --fail-on-regression behavior pass-rate-only for backward compatibility, while adding explicit metric gate flags for latency, p95, cost, and any regression
• persist p95_latency_ms in recorded history and ignore malformed/non-run JSON files in history directories so trend output does not poison later analyses
• document the new trend workflow and add focused CLI/history coverage

Why now

• Teams are asking for better production eval drift detection than single-run pass/fail checks: https://www.reddit.com/r/AI_Agents/comments/1s7n02g/what_are_the_best_methods_to_evaluate_the/
• Reliability and observability for agentic systems is a live infra wedge, not just a research problem: https://www.arthur.ai/column/agentic-ai-observability-playbook-2026
• Launch momentum is clustering around AI agent and AI infrastructure tooling, which makes release-gating UX a strong PM-facing surface: https://www.producthunt.com/categories/ai-infrastructure

Validation

• python3 -m pytest tests/test_history.py tests/test_cli.py ✅ (16 passed)
• python3 -m pytest ✅ (27 passed)
• python3 -m pip install . ✅ (wheel/install succeeded locally; host pip 21.2.4 emitted legacy warnings)
• python3 -m compileall src ✅
• PYTHONPATH=src python3 -m agent_release_gate.cli evaluate --spec examples/spec.yaml --results examples/results.json --output "$TMP_DIR/gate-report.json" --markdown "$TMP_DIR/gate-report.md" --record-history "$TMP_DIR/history" ✅ smoke run, gate passed, recorded p95_latency_ms=920.0
• PYTHONPATH=src python3 -m agent_release_gate.cli trend --history "$TMP_DIR" --fail-on-latency-regression > "$TMP_DIR-output.json" ✅ smoke run, exited 1 with avg_latency_regression=true, p95_latency_regression=true, avg_cost_regression=true
• node /Users/bwise/.openclaw/workspace/projects/humanizer/src/cli.js analyze -f README.md --ignore-code --json ✅ humanizer score 1, high-confidence, no significant AI-writing patterns detected

Risk / rollback

• Low risk, additive CLI surface only
• Roll back by reverting commit 5022edf