fix(exec): strip trailing result-schema JSON from local model output

Small models (phi4-mini) produce correct markdown analysis but then
append the old {status/phase/skill} JSON schema out of training habit.
stripResultJSON() detects and removes these trailing fences so Claude
Code receives clean prose regardless of model behaviour.

Non-schema json blocks (config examples etc) are preserved.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

internal/exec/litellm.go

@@ -6,6 +6,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"net/http"
+	"strings"
 	"time"
 )
 
@@ -93,5 +94,34 @@ func (e *LiteLLMExecutor) Complete(ctx context.Context, model, system, user stri
 		return "", 0, fmt.Errorf("litellm: no choices in response")
 	}
 
-	return chatResp.Choices[0].Message.Content, durationMs, nil
+	return stripResultJSON(chatResp.Choices[0].Message.Content), durationMs, nil
+}
+
+// stripResultJSON removes trailing ```json blocks that match the old structured
+// result schema (containing "status" and "phase" keys). Some local models produce
+// correct markdown prose but then append the old JSON format out of habit.
+func stripResultJSON(text string) string {
+	const fence = "```json"
+	idx := len(text) - 1
+	// Walk backwards past trailing whitespace.
+	for idx >= 0 && (text[idx] == '\n' || text[idx] == '\r' || text[idx] == ' ') {
+		idx--
+	}
+	// Must end with closing fence.
+	if idx < 2 || text[idx-2:idx+1] != "```" {
+		return text
+	}
+	// Find the matching opening fence.
+	start := len(text[:idx-2]) - 1
+	for start >= 0 {
+		if start+len(fence) <= len(text) && text[start:start+len(fence)] == fence {
+			block := text[start : idx+1]
+			if strings.Contains(block, `"status"`) && strings.Contains(block, `"phase"`) {
+				return strings.TrimRight(text[:start], " \t\r\n")
+			}
+			break
+		}
+		start--
+	}
+	return text
 }