fix(tts/kokoro-ane): adopt COLA-corrected KokoroTail_v2 + native output level for all variants

Sources/FluidAudio/ModelNames.swift

@@ -1428,7 +1428,10 @@ public enum ModelNames {
         // re-download. See mobius laishere-coreml docs/trials-and-errors.md.
         public static let noise = "KokoroNoise_v2.mlmodelc"
         public static let vocoder = "KokoroVocoder.mlmodelc"
-        public static let tail = "KokoroTail.mlmodelc"
+        // v2: COLA-normalized iSTFT deconv weights (raw output was exactly 1.5x
+        // the PyTorch reference). Renamed (not overwritten) so cached clients
+        // re-download. See issue #852.
+        public static let tail = "KokoroTail_v2.mlmodelc"
 
         /// Auxiliary (non-CoreML) files that must accompany the mlmodelc bundles.
         public static let vocab = "vocab.json"

Sources/FluidAudio/TTS/KokoroAne/KokoroAneConstants.swift

@@ -16,7 +16,7 @@ public enum KokoroAneConstants {
     /// Default voice id for the Japanese (`ANE-ja/`) variant.
     public static let defaultVoiceJapanese = "jf_alpha"
 
-    /// Output sample rate of the iSTFT in `KokoroTail.mlpackage`.
+    /// Output sample rate of the iSTFT in `KokoroTail_v2.mlpackage`.
     public static let sampleRate = 24_000
 
     /// BOS / EOS token id used by both `convert-coreml.py` and the iOS demo.

Sources/FluidAudio/TTS/KokoroAne/KokoroAneManager.swift

@@ -358,14 +358,14 @@ public actor KokoroAneManager {
 
     private func wavData(from result: KokoroAneSynthesisResult) throws -> Data {
         do {
-            // Japanese writes at the model's native level (no peak-normalization)
-            // so the output matches the PyTorch reference instead of being
-            // slammed to 0 dBFS. English/Mandarin keep peak-normalization until
-            // their tails get the same COLA-corrected iSTFT (#698 follow-up).
+            // All variants write at the model's native level (no
+            // peak-normalization) so the output matches the PyTorch reference
+            // instead of being slammed to 0 dBFS. Requires the COLA-corrected
+            // KokoroTail_v2 (#852).
             return try AudioWAV.data(
                 from: result.samples,
                 sampleRate: Double(result.sampleRate),
-                normalize: variant != .japanese)
+                normalize: false)
         } catch {
             throw KokoroAneError.audioConversionFailed(error.localizedDescription)
         }

Sources/FluidAudio/TTS/KokoroAne/Pipeline/KokoroAneSynthesizer+Types.swift

@@ -96,7 +96,7 @@ public enum KokoroAneStage: String, CaseIterable, Sendable {
         case .prosody: return "KokoroProsody.mlmodelc"
         case .noise: return "KokoroNoise_v2.mlmodelc"  // v2: atan2 phase-correction (HF-noise fix)
         case .vocoder: return "KokoroVocoder.mlmodelc"
-        case .tail: return "KokoroTail.mlmodelc"
+        case .tail: return "KokoroTail_v2.mlmodelc"  // v2: COLA-normalized iSTFT (level fix, #852)
         }
     }
 }

Sources/FluidAudioCLI/Commands/TTSCommand.swift

@@ -879,10 +879,12 @@ public struct TTS {
                 detailed = try await manager.synthesizeDetailed(
                     text: text, voice: resolvedVoice, speed: 1.0)
             }
+            // Native level for all variants — matches the PyTorch reference
+            // now that KokoroTail_v2 carries the COLA-corrected iSTFT (#852).
             let wav = try AudioWAV.data(
                 from: detailed.samples,
                 sampleRate: Double(detailed.sampleRate),
-                normalize: variant != .japanese)
+                normalize: false)
             let tSynth1 = Date()
 
             let outURL = resolveInputURL(output)