test: added live openai tests

doc/test-plan.md

@@ -1341,6 +1341,7 @@ _Document results here as tests are completed_
 | 2.3.x   | ✅ PASSED | 2026-04-07 | OpenCode | 2-minute timeout behavior and network failure handling validated |
 | 2.4.x   | ✅ PASSED | 2026-04-07 | OpenCode | JSON parsing success/failure/empty choices/missing fields covered |
 | 2.5.x   | ✅ PASSED | 2026-04-07 | OpenCode | Upstream 500/429 and malformed content error handling validated |
+| 2.x live consistency | ⏭️ READY (manual run) | 2026-04-07 | OpenCode | Added opt-in live test `TestCallLLM_Live_ScoreConsistencyPlusMinus10` gated by `RUN_LIVE_OPENAI_TESTS=1` |
 | 3.1.x   | ✅ PASSED | 2026-04-07 | OpenCode | Overall score range and consistency scenarios covered with deterministic mocks |
 | 3.2.x   | ✅ PASSED | 2026-04-07 | OpenCode | Criteria population, score bounds, evidence/comments presence verified |
 | 3.3.x   | ✅ PASSED | 2026-04-07 | OpenCode | Strengths/weaknesses/missing information structure checks validated |
@@ -1397,7 +1398,7 @@ _Document any test failures here with details_
 
 **Next Steps:**
 1. Add CI step to run `go test ./...` before image build/push
-2. Tighten manual validation notes for production-like OpenAI calls
+2. Run opt-in live consistency test (`RUN_LIVE_OPENAI_TESTS=1`) and log observed score distribution
 3. Backfill health-check endpoint or document non-AI endpoint strategy
 4. Keep the intentional 1.2.6 skip documented until encrypted-PDF fixtures are added
 

internal/services/analyzer_live_test.go

@@ -0,0 +1,77 @@
+package services
+
+import (
+	"math"
+	"os"
+	"testing"
+	"time"
+)
+
+// TestCallLLM_Live_ScoreConsistencyPlusMinus10 validates SRD_NonFuncReq_0006 /
+// SRD_QualAssurReq_0001 against the live OpenAI API.
+//
+// This test is opt-in to avoid API cost/flakiness in default runs:
+// RUN_LIVE_OPENAI_TESTS=1 OPENAI_API_KEY=... go test ./internal/services -run TestCallLLM_Live_ScoreConsistencyPlusMinus10 -v
+func TestCallLLM_Live_ScoreConsistencyPlusMinus10(t *testing.T) {
+	if os.Getenv("RUN_LIVE_OPENAI_TESTS") != "1" {
+		t.Skip("set RUN_LIVE_OPENAI_TESTS=1 to run live OpenAI consistency test")
+	}
+	if os.Getenv("OPENAI_API_KEY") == "" {
+		t.Skip("OPENAI_API_KEY is required for live OpenAI test")
+	}
+
+	resume := `Senior Software Engineer with 7 years of experience building Go backend services.
+Led microservice migrations, improved API latency by 35%, and maintained CI/CD pipelines.
+Experience includes Kubernetes, Docker, PostgreSQL, and cloud deployments on AWS.`
+
+	job := `We are hiring a Senior Go Backend Engineer with strong API design skills,
+production Kubernetes experience, and ownership of scalable distributed systems.
+Candidates should demonstrate measurable impact, collaboration, and code quality.`
+
+	const runs = 10
+	scores := make([]int, 0, runs)
+
+	for i := range runs {
+		result, err := callLLM(resume, job)
+		if err != nil {
+			t.Fatalf("run %d failed: %v", i+1, err)
+		}
+		if result.OverallScore < 0 || result.OverallScore > 100 {
+			t.Fatalf("run %d produced out-of-range score: %d", i+1, result.OverallScore)
+		}
+
+		scores = append(scores, result.OverallScore)
+		t.Logf("run %d score: %d", i+1, result.OverallScore)
+
+		if i < runs-1 {
+			time.Sleep(300 * time.Millisecond)
+		}
+	}
+
+	baseline := scores[0]
+	for i, score := range scores {
+		delta := score - baseline
+		if delta < 0 {
+			delta = -delta
+		}
+		if delta > 10 {
+			t.Fatalf("run %d score %d exceeded +/-10 bound from baseline %d", i+1, score, baseline)
+		}
+	}
+
+	var sum float64
+	for _, s := range scores {
+		sum += float64(s)
+	}
+	mean := sum / float64(len(scores))
+
+	var variance float64
+	for _, s := range scores {
+		d := float64(s) - mean
+		variance += d * d
+	}
+	variance /= float64(len(scores))
+	stddev := math.Sqrt(variance)
+
+	t.Logf("baseline=%d scores=%v mean=%.2f stddev=%.2f", baseline, scores, mean, stddev)
+}