Stabilize narration voice and simplify video bookends

macos/Astroshots/Narration/NarrationModelManager.swift

@@ -305,12 +305,15 @@ final class NarrationSpeechBox: @unchecked Sendable {
 
     func speak(_ text: String, voice: String = NarrationDefaults.voice) async throws -> [Float] {
         let trimmed = text.trimmingCharacters(in: .whitespacesAndNewlines)
+        var generationParameters = model.defaultGenerationParameters
+        generationParameters.temperature = NarrationDefaults.temperature
         let audio: MLXArray = try await model.generate(
             text: trimmed,
             voice: NarrationVoice.normalized(voice),
             refAudio: nil,
             refText: nil,
-            language: NarrationDefaults.language
+            language: NarrationDefaults.language,
+            generationParameters: generationParameters
         )
         return audio.asArray(Float.self)
     }

macos/Astroshots/Narration/NarrationModels.swift

@@ -6,6 +6,8 @@ enum NarrationDefaults {
     /// CustomVoice English preset (see Qwen3-TTS README).
     static let voice = "Ryan"
     static let language = "English"
+    /// Keep independently synthesized friction-log steps sounding like one narrator.
+    static let temperature: Float = 0.3
     /// Minimum step display duration when audio is shorter.
     static let minimumStepSeconds: Double = 1.6
     static let framesPerSecond: Int32 = 12

macos/Astroshots/Narration/NarrationVideoComposer.swift

@@ -143,9 +143,7 @@ enum NarrationVideoComposer {
         }
         writer.startSession(atSourceTime: .zero)
 
-        let image = step.isBrandCard
-            ? (brandBackground() ?? placeholderImage(size: size))
-            : (loadImage(path: step.imagePath) ?? placeholderImage(size: size))
+        let image = loadImage(path: step.imagePath) ?? placeholderImage(size: size)
         // Still-image segments stay inexpensive while fades remain visibly smooth.
         let fps = NarrationDefaults.framesPerSecond
         let frameCount = max(2, Int(ceil(duration * Double(fps))))
@@ -434,11 +432,13 @@ enum NarrationVideoComposer {
                 | CGBitmapInfo.byteOrder32Little.rawValue
         ) else { return nil }
 
-        context.setFillColor(NSColor(calibratedWhite: 0.035, alpha: 1).cgColor)
+        let backgroundColor = step.isBrandCard
+            ? NSColor.black
+            : NSColor(calibratedWhite: 0.035, alpha: 1)
+        context.setFillColor(backgroundColor.cgColor)
         context.fill(CGRect(origin: .zero, size: size))
 
         if step.isBrandCard {
-            drawAspectFill(image, in: CGRect(origin: .zero, size: size), context: context)
             drawBrandOverlay(title: step.title, size: size, context: context)
             return buffer
         }
@@ -512,14 +512,6 @@ enum NarrationVideoComposer {
         return chunks[min(Int(progress * Double(chunks.count)), chunks.count - 1)]
     }
 
-    private static func brandBackground() -> NSImage? {
-        guard let url = Bundle.main.url(
-            forResource: "astroshots-dmg-background",
-            withExtension: "png"
-        ) else { return nil }
-        return NSImage(contentsOf: url)
-    }
-
     private static func brandIcon() -> NSImage? {
         guard let url = Bundle.main.url(
             forResource: "astroshots-app-icon-master",
@@ -549,32 +541,6 @@ enum NarrationVideoComposer {
         context.draw(cgImage, in: rect)
     }
 
-    private static func drawAspectFill(
-        _ image: NSImage,
-        in frame: CGRect,
-        context: CGContext
-    ) {
-        guard let cgImage = image.cgImage(forProposedRect: nil, context: nil, hints: nil) else {
-            return
-        }
-        let imageSize = CGSize(width: cgImage.width, height: cgImage.height)
-        let scale = max(frame.width / imageSize.width, frame.height / imageSize.height)
-        let drawSize = CGSize(width: imageSize.width * scale, height: imageSize.height * scale)
-        context.saveGState()
-        context.clip(to: frame)
-        context.interpolationQuality = .high
-        context.draw(
-            cgImage,
-            in: CGRect(
-                x: frame.midX - drawSize.width / 2,
-                y: frame.midY - drawSize.height / 2,
-                width: drawSize.width,
-                height: drawSize.height
-            )
-        )
-        context.restoreGState()
-    }
-
     private static func withFlippedAppKitContext(
         _ context: CGContext,
         draw: () -> Void

macos/AstroshotsTests/NarrationTests.swift

@@ -62,6 +62,7 @@ struct NarrationTests {
     @Test func defaultModelIsQwen3TTS() {
         #expect(NarrationDefaults.modelID.contains("Qwen3-TTS"))
         #expect(NarrationDefaults.voice == "Ryan")
+        #expect(NarrationDefaults.temperature == 0.3)
         #expect(NarrationVoice.available.map(\.id).contains("Aiden"))
     }
 
@@ -129,7 +130,11 @@ struct NarrationTests {
                 kCVPixelBufferPixelFormatTypeKey as String: Int(kCVPixelFormatType_32BGRA),
             ]
         )
-        #expect(CMSampleBufferGetImageBuffer(videoSample) != nil)
+        let titleFrame = try #require(CMSampleBufferGetImageBuffer(videoSample))
+        let titleCorner = try pixelRGB(in: titleFrame, x: 8, y: 8)
+        #expect(titleCorner.red < 16)
+        #expect(titleCorner.green < 16)
+        #expect(titleCorner.blue < 16)
 
         let audioSample = try decodeFirstSample(
             from: asset,
@@ -311,4 +316,17 @@ struct NarrationTests {
         }
         return sample
     }
+
+    private func pixelRGB(
+        in buffer: CVPixelBuffer,
+        x: Int,
+        y: Int
+    ) throws -> (red: UInt8, green: UInt8, blue: UInt8) {
+        CVPixelBufferLockBaseAddress(buffer, .readOnly)
+        defer { CVPixelBufferUnlockBaseAddress(buffer, .readOnly) }
+        let base = try #require(CVPixelBufferGetBaseAddress(buffer))
+        let offset = y * CVPixelBufferGetBytesPerRow(buffer) + x * 4
+        let pixel = base.advanced(by: offset).assumingMemoryBound(to: UInt8.self)
+        return (red: pixel[2], green: pixel[1], blue: pixel[0])
+    }
 }