feat: phase 6 — metric catalogue + threshold scoring

[barancan]

Summary

• Metric catalogue replaces the 7 starter chips with 12 categorised entries (format, correctness, safety, performance, rag, quality), each carrying determinism, default_threshold, and the requirement text used to toggle it
• Threshold scoring — CheckResult gains a threshold field; from_raw() overrides passed = score >= threshold when both are present; threshold flows through to CaseResult.checks JSON for display in transcripts
• Generator — keyword detection expanded to cover toxicity (≥0.90), hallucination (≥0.85), faithfulness / answer relevance / task completion (≥0.80); thresholds embedded into check with dict at generation time so they're editable via the existing PATCH endpoint
• Wizard Step 1 — chips grouped by category with section headers; active chip colour matches determinism (blue = deterministic, purple = stochastic); stochastic chips show (≥0.80) threshold hint
• Wizard Step 2 preview — judge check rows show threshold next to the assertion
• Transcript — score vs threshold display (0.72 < 0.80) was stubbed in Phase 5 and is now live

Test plan

• pytest -q → 149 passed
• Open /pipelines/new — Step 1 chips should be grouped by category (format, safety, rag, etc.) with section labels; stochastic chips should be purple when active and show threshold hints
• Select "RAG faithfulness" chip — requirement text should appear in the textarea
• Generate a pipeline with a toxicity/faithfulness requirement — confirm the generated check with dict contains threshold on the review page
• Run a pipeline with a graded judge check — confirm transcript shows score ≥ threshold or score < threshold format
• Verify deterministic chips (Valid JSON, No PII) show no threshold hint

🤖 Generated with Claude Code