fix: remove dead audit/graph surface, surface LLM errors, FailOn contract

.github/workflows/ci.yml

@@ -32,7 +32,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  GO_VERSION: "1.26.5"
+  GO_VERSION: "1.26.6"
   GOPROXY: "https://proxy.golang.org,direct"
   GOPRIVATE: "github.com/GrayCodeAI/*"
   GONOSUMDB: "github.com/GrayCodeAI/*"

CHANGELOG.md

@@ -7,6 +7,40 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 ---
 
+## [Unreleased]
+
+### Changed
+- Extracted the out-of-range confidence fallback in `toPublicFindings`
+  into a documented package constant, `defaultConfidence` (0.6). No
+  behavior change.
+
+### Fixed
+- **Configured fail thresholds now survive contract conversion.**
+  `ToContractResult` copied `FailOn` by field assignment, leaving the
+  contract's `FailOnSet` false so its `Failed()` ignored a
+  user-configured below-critical threshold (e.g. `WithFailOn(High)`
+  from the `CI` preset). Conversion now records the threshold via
+  `SetFailOn`, so the contract `Result.Failed()` honors it.
+- **Provider failures are now visible in `Stats`.** A review in which
+  every LLM call failed previously succeeded silently — errors only
+  appeared in the human-readable `Report`. `Stats.LLMErrors` now
+  records one entry per failed call (concern calls and the
+  self-reflection pass), and `ToContractResult` carries them into the
+  shared `reviewcontracts.Stats.LLMErrors` field so hawk can surface
+  partial results.
+
+### Removed
+- **Dead audit/graph API surface** (breaking, pre-1.0). The
+  `WithGraph`, `WithAuditMode`, and `WithAuditTargets` options and the
+  `AuditMode`, `AuditTarget`, `AuditTargetType`, `AuditOption`, and
+  `ParseAuditMode` types configured `Reviewer` fields that were never
+  read, and the `internal/graph` and `internal/audit` packages had no
+  callers. Removed the options, types, packages, the unused `Reviewer`
+  fields, and the `graph`/`audit` `.sight.toml` keys (the keys were
+  silently ignored in effect; they now remain unparsed, which is the
+  same behavior). Consumers should use `qualitygraph` for graph
+  projection and `inspect` for deployed-surface auditing.
+
 ## [0.1.2] - 2026-07-04
 
 ### Changed

config.go

@@ -19,8 +19,6 @@ type FileConfig struct {
 	Reflection *bool             `json:"reflection"`
 	Parallel   *bool             `json:"parallel"`
 	Prompts    map[string]string `json:"prompts"`
-	Graph      *bool             `json:"graph"`
-	Audit      *string           `json:"audit"` // "none", "hooks", "mcp", "full"
 }
 
 // LoadConfigFile reads .sight.toml from the given directory (or parents).
@@ -80,12 +78,6 @@ func ApplyFileConfig(fc *FileConfig) []Option {
 	if len(fc.Exclude) > 0 {
 		opts = append(opts, WithExclude(fc.Exclude...))
 	}
-	if fc.Graph != nil {
-		opts = append(opts, WithGraph(*fc.Graph))
-	}
-	if fc.Audit != nil {
-		opts = append(opts, WithAuditMode(ParseAuditMode(*fc.Audit)))
-	}
 
 	return opts
 }
@@ -160,11 +152,6 @@ func parseTOMLConfig(content string) (*FileConfig, error) {
 		case "parallel":
 			b := value == "true"
 			cfg.Parallel = &b
-		case "graph":
-			b := value == "true"
-			cfg.Graph = &b
-		case "audit":
-			cfg.Audit = &value
 		}
 	}
 

contracts.go

@@ -94,6 +94,7 @@ func toContractStats(s Stats) reviewcontracts.Stats {
 		AverageConfidence:   s.AverageConfidence,
 		HighConfidenceCount: s.HighConfidenceCount,
 		LowConfidenceCount:  s.LowConfidenceCount,
+		LLMErrors:           s.LLMErrors,
 	}
 }
 
@@ -113,13 +114,17 @@ func ToContractResult(r *Result) *reviewcontracts.Result {
 	if r == nil {
 		return nil
 	}
-	return &reviewcontracts.Result{
+	res := &reviewcontracts.Result{
 		Findings:            ToContractFindings(r.Findings),
 		Comments:            ToContractInlineComments(r.Comments),
 		Stats:               toContractStats(r.Stats),
 		Report:              r.Report,
-		FailOn:              r.FailOn,
 		SASTFusion:          toContractSASTFusion(r.SASTFusion),
 		ConfidenceBreakdown: toContractConfidenceBreakdown(r.ConfidenceBreakdown),
 	}
+	// Set the threshold through SetFailOn so FailOnSet is recorded: the
+	// contract's Failed() ignores a directly-assigned FailOn below
+	// critical, which would drop a user-configured threshold.
+	res.SetFailOn(r.FailOn)
+	return res
 }

contracts_test.go

@@ -1,6 +1,9 @@
 package sight
 
-import "testing"
+import (
+	"context"
+	"testing"
+)
 
 func TestToContractResult(t *testing.T) {
 	t.Parallel()
@@ -49,3 +52,65 @@ func TestToContractResult(t *testing.T) {
 		t.Fatal("expected confidence breakdown to convert")
 	}
 }
+
+func TestToContractResult_FailOnThresholdTakesEffect(t *testing.T) {
+	t.Parallel()
+
+	// A below-critical threshold configured on the sight Result must
+	// survive conversion: the contract's Failed() honors it only when
+	// FailOnSet is true, which ToContractResult must arrange via SetFailOn.
+	result := &Result{
+		FailOn: SeverityHigh,
+		Findings: []Finding{
+			{Severity: SeverityInfo, Message: "note", Confidence: 0.5},
+		},
+	}
+
+	contract := ToContractResult(result)
+	if !contract.FailOnSet {
+		t.Fatal("FailOnSet = false, want true after conversion")
+	}
+	if contract.FailOn != SeverityHigh {
+		t.Fatalf("FailOn = %v, want high", contract.FailOn)
+	}
+	if contract.Failed() {
+		t.Error("info finding must not fail a review with a high threshold")
+	}
+
+	result.Findings = append(result.Findings, Finding{
+		Severity: SeverityHigh, Message: "real problem", Confidence: 0.5,
+	})
+	contract = ToContractResult(result)
+	if !contract.Failed() {
+		t.Error("high finding must fail a review with a high threshold")
+	}
+}
+
+func TestToContractResult_FailOnConfiguredViaOptions(t *testing.T) {
+	t.Parallel()
+
+	reviewWith := func(response string) *Result {
+		t.Helper()
+		r := NewReviewer(
+			WithProvider(&fixMockProvider{response: response}),
+			WithFailOn(SeverityHigh),
+			WithConcerns("security"),
+			WithParallel(false),
+		)
+		result, err := r.Review(context.Background(), sampleDiff)
+		if err != nil {
+			t.Fatalf("Review failed: %v", err)
+		}
+		return result
+	}
+
+	infoOnly := reviewWith(`[{"file": "handler.go", "line": 13, "severity": "info", "message": "style nit", "fix": "n/a"}]`)
+	if got := ToContractResult(infoOnly).Failed(); got {
+		t.Error("Failed() = true for info-only findings with failOn=high, want false")
+	}
+
+	highFinding := reviewWith(`[{"file": "handler.go", "line": 13, "severity": "high", "message": "SQL injection", "fix": "use params"}]`)
+	if got := ToContractResult(highFinding).Failed(); !got {
+		t.Error("Failed() = false for a high finding with failOn=high, want true")
+	}
+}

docs/architecture.md

@@ -134,63 +134,6 @@ srv.ServeHTTP("127.0.0.1:8080") // 🌐 streamable HTTP transport, served at /mc
 
 ---
 
-## 🔗 Structural Dependency Graph
-
-The `internal/graph` package provides **optional structural dependency analysis** for code reviews:
-
-**Key capabilities:**
-- Parse Go AST to build structural dependency graph
-- Blast-radius analysis: identify all files affected by changes
-- Transitive dependency tracking
-- Impact scoring based on depth and number of dependents
-- SQLite persistence for incremental updates
-
-**Usage:**
-```go
-// Enable graph-backed review
-g := graph.New()
-// ... build graph or load from SQLite ...
-
-// Run blast radius analysis
-result := g.GetBlastRadius([]string{"file.go"})
-fmt.Printf("Direct: %d, Transitive: %d\n", result.Direct, result.Transitive)
-```
-
-**Tools exposed via MCP:** `sight_graph_blastRadius`, `sight_graph_query`, `sight_graph_stats`
-
----
-
-## 🔒 Security Auditing
-
-The `internal/audit` package provides **agent surface security auditing**:
-
-**Key capabilities:**
-- Concurrent security scanning of multiple targets
-- Filter findings by severity (critical, high, medium, low) or category
-- Integration with MCP hooks, permissions, and endpoints
-- Webhook validation and test requests
-- JSON-serializable audit findings
-
-**Usage:**
-```go
-// Create audit scope
-auditScope := &audit.AuditScope{
-    Targets: []audit.AuditTarget{
-        {Type: audit.AuditTargetHooks, Path: "/hooks/webhook"},
-        {Type: audit.AuditTargetEndpoints, Path: "/api/endpoint"},
-    },
-    Rules: []string{"detect_unauthenticated_endpoints", "detect_hardcoded_secrets"},
-}
-
-// Run audit
-report, err := audit.Audit(ctx, auditScope)
-if report.Count() > 0 {
-    fmt.Printf("Found %d security issues\n", report.Count())
-}
-```
-
----
-
 ## 🌐 Browser Automation Tools
 
 The `internal/tool` package provides **HTTP-based browser automation** capabilities: