From 058d30bfd0de2ead0d72a4324dd815ba320ca846 Mon Sep 17 00:00:00 2001
From: Sanchit Monga <sanchitmonga22@gmail.com>
Date: Wed, 2 Sep 2026 02:40:49 -0700
Subject: [PATCH] feat(catalog): Kokoro-82M on the ANE, and pin SDK 0.20.34

The app's first ANE text-to-speech row, and the last of NeuRT's eight served
primitives to reach a catalog. Kokoro-82M runs across three Core ML graphs
(duration -> decode -> gen) plus two host seams that are not expressible as ANE
ops -- the duration->alignment expansion and the harmonic source -- and ships
its own G2P lexicon, so no runtime phonemizer is needed.

0.20.34, not 0.20.33, and the distinction matters. The tts_ops slot shipped in
.33, but the engine emitted int16 PCM where every other engine emits float32
under the same AUDIO_FORMAT_PCM enum -- which names a CONTAINER and carries no
bit depth at all. Consumers reading the documented float32 got half the samples
and a saturated waveform, with no error at any layer. Fixed in the 0.20.34
engine, so a row pinned to .33 would have downloaded 146 MB and played garbage.

`.tts` sets allowedFrameworks nil, so this row is reachable with no picker
change -- unlike the RAG contexts, which had to gain `.coreml` explicitly.

    swift package resolve      runanywhere-swift 0.20.34 (8bd75a29)
    xcodebuild test            ** TEST SUCCEEDED **  23 passed, 0 failed

Verified end to end on the same engine through RCLI, which shares this SDK:
79200 samples, 3.30s @ 24 kHz, 0.00% clipped.
---
 Package.resolved                              |  6 ++---
 Package.swift                                 |  2 +-
 RunAnywhereAI.xcodeproj/project.pbxproj       |  2 +-
 .../Core/Services/ModelCatalogBootstrap.swift | 24 +++++++++++++++++++
 4 files changed, 29 insertions(+), 5 deletions(-)

diff --git a/Package.resolved b/Package.resolved
index eebfd6e..30d7151 100644
--- a/Package.resolved
+++ b/Package.resolved
@@ -1,5 +1,5 @@
 {
-  "originHash" : "8c8d876c8a7cb371cf2bcef775986d0f6b40d5aaff481da000c5fca5c9e4707d",
+  "originHash" : "265319a3e745dcc48d7923b465514f940d8689aae2a78ddbe2a19fef1f415f9a",
   "pins" : [
     {
       "identity" : "devicekit",
@@ -60,8 +60,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/RunanywhereAI/runanywhere-swift.git",
       "state" : {
-        "revision" : "4d1a0480eb80f516da481607fb86a7ad3f566432",
-        "version" : "0.20.33"
+        "revision" : "8bd75a291dee4c359a2c118ee317ab8c33dd0238",
+        "version" : "0.20.34"
       }
     },
     {
diff --git a/Package.swift b/Package.swift
index c664122..ac653a4 100644
--- a/Package.swift
+++ b/Package.swift
@@ -63,7 +63,7 @@ let package = Package(
         // needs.
         .package(
             url: "https://github.com/RunanywhereAI/runanywhere-swift.git",
-            from: "0.20.33"
+            from: "0.20.34"
         ),
     ],
     targets: [
diff --git a/RunAnywhereAI.xcodeproj/project.pbxproj b/RunAnywhereAI.xcodeproj/project.pbxproj
index 0aac1e5..c02d220 100644
--- a/RunAnywhereAI.xcodeproj/project.pbxproj
+++ b/RunAnywhereAI.xcodeproj/project.pbxproj
@@ -1031,7 +1031,7 @@
 			repositoryURL = "https://github.com/RunanywhereAI/runanywhere-swift.git";
 			requirement = {
 				kind = upToNextMajorVersion;
-				minimumVersion = 0.20.33;
+				minimumVersion = 0.20.34;
 			};
 		};
 /* End XCRemoteSwiftPackageReference section */
diff --git a/RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift b/RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift
index bd75f30..979c6cc 100644
--- a/RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift
+++ b/RunAnywhereAI/Core/Services/ModelCatalogBootstrap.swift
@@ -1599,6 +1599,30 @@ enum ModelCatalogBootstrap {
         logger.info("LoRA adapters skipped: no adapter matches the current catalog")
         #endif
 
+        // The first ANE TEXT-TO-SPEECH row, and NeuRT's last null primitive filled. Kokoro-82M
+        // across three Core ML graphs (duration -> decode -> gen) plus two host seams that are not
+        // expressible as ANE ops: the duration->alignment expansion and the harmonic source. Ships
+        // its own G2P lexicon, so no runtime phonemizer is needed.
+        //
+        // Requires SDK 0.20.34, not 0.20.33. The slot shipped in .33, but the engine emitted int16
+        // PCM where every other engine emits float32 under the same AUDIO_FORMAT_PCM enum -- which
+        // names a container and carries no bit depth -- so audio came out half-length and
+        // saturated with no error anywhere. Fixed in the 0.20.34 engine.
+        //
+        // `.tts` sets allowedFrameworks nil, so this row is reachable without a picker change.
+        await registerArchive(
+            id: "kokoro-82m-ane",
+            name: "Kokoro 82M (NeuRT / Neural Engine)",
+            url: "https://huggingface.co/runanywhere/Kokoro-82M_ANE/resolve/main/"
+                + "kokoro-82m_ANE.zip",
+            framework: .coreml,
+            modality: .speechSynthesis,
+            archive: .zip,
+            structure: .nestedDirectory,
+            // 145.9 MB download; the three graphs plus a 4.3 MB G2P lexicon resident.
+            memoryRequirement: 400_000_000
+        )
+
         // --- Image generation (CoreML diffusion; Apple only) --------------------
         // Apple's palettized Stable Diffusion 1.5. The id matches the built-in
         // diffusion registry (diffusion_model_registry.cpp) and RCLI's `sd15`