On-device fixes: optimized native build, 16 KB align, disable thinking

Validated on a Pixel 7 (adb). Key fixes:
- Force Release/-O3 for native even in the debug variant (AGP defaulted to -O0,
  making llama.cpp ~10x too slow / effectively unusable).
- 16 KB-align all native LOAD segments (Play requirement; 16 KB-page devices).
- Disable Qwen3 thinking (empty <think></think> prefill + /no_think) so replies
  are fast and punchy instead of spending the whole budget on hidden reasoning.
- Cap replies at 220 tokens.

Verified: model downloads, loads (n_ctx=4096, 6 threads), streams an on-persona
reply with emoji intact, and "new tsjet" clears the conversation.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -34,7 +34,9 @@ conversation and start fresh.
   **downloaded on first launch** from Hugging Face into the app's private storage.
   It is *not* bundled in the APK. (Swap the quant in `ModelDownloader.kt`.)
 - **Persona:** set via `SYSTEM_PROMPT` in `MainActivity.kt`; sampling is a little
-  hot (temp 0.9) for playful answers.
+  hot (temp 0.9) for playful answers. The model is a Qwen3 "thinking" model, so the
+  prompt disables reasoning (empty `<think></think>` prefill + `/no_think`) to keep
+  replies fast and punchy instead of burning the token budget on hidden thoughts.
 - **Conversation:** the web layer holds the full history and sends it each turn;
   native rebuilds the ChatML prompt and clears the KV cache before every reply, so
   "new tsjet" is just: clear JS state + reset cache.
@@ -85,5 +87,9 @@ manager, and tap it. First launch downloads the ~1.9 GB model over Wi-Fi.
   display and from history (`stripThink` in `app.js`).
 - Bumping the llama.cpp tag? Re-check the C API calls in `llama-jni.cpp` against
   that tag's `include/llama.h` — it uses the raw C API directly.
-- **Build verified:** `./gradlew assembleDebug` produces a ~13 MB `arm64-v8a` APK
+- **Native is always built optimized.** AGP compiles the debug variant's C/C++ at
-  (native libs stripped; the ~1.9 GB model downloads on first launch).
+  `-O0` by default, which makes llama.cpp ~10x too slow; `CMakeLists.txt` forces
+  `Release`/`-O3` regardless of variant. Native libs are also 16 KB page-aligned.
+- **Verified on-device** (Pixel 7, Android, 4 KB pages): downloads the model, loads
+  it (`n_ctx=4096`, 6 threads), streams a reply, and "new tsjet" clears the chat.
+  `./gradlew assembleDebug` → ~13 MB `arm64-v8a` APK (model downloads on first launch).

app/build.gradle.kts

@@ -24,7 +24,10 @@ android {
         }
         externalNativeBuild {
             cmake {
+                // Force an optimized native build even for the debug APK — otherwise
+                // AGP compiles ggml/llama at -O0 and inference is ~10x too slow.
                 arguments += "-DANDROID_STL=c++_shared"
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
                 cppFlags += "-std=c++17"
             }
         }

app/src/main/cpp/CMakeLists.txt

@@ -2,6 +2,19 @@ cmake_minimum_required(VERSION 3.22.1)
 
 project(tsjet LANGUAGES C CXX)
 
+# AGP compiles the debug variant with CMAKE_BUILD_TYPE=Debug (-O0), which makes
+# llama.cpp ~10x too slow. Force an optimized build regardless of the variant.
+if(NOT CMAKE_BUILD_TYPE OR CMAKE_BUILD_TYPE STREQUAL "Debug")
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "" FORCE)
+endif()
+set(CMAKE_C_FLAGS_DEBUG   "-O3 -DNDEBUG")
+set(CMAKE_CXX_FLAGS_DEBUG "-O3 -DNDEBUG")
+
+# Align all native LOAD segments to 16 KB. Required for Android 15+ / devices with
+# 16 KB memory pages and for Google Play uploads. Set BEFORE llama.cpp is added so
+# its targets (libllama, libggml*) inherit it too.
+add_link_options("-Wl,-z,max-page-size=16384" "-Wl,-z,common-page-size=16384")
+
 include(FetchContent)
 
 # --- llama.cpp (pinned) -----------------------------------------------------

app/src/main/java/monster/autisme/tsjetpiti/MainActivity.kt

@@ -34,7 +34,7 @@ class MainActivity : Activity() {
 
     companion object {
         private const val N_CTX = 4096
-        private const val MAX_TOKENS = 384
+        private const val MAX_TOKENS = 220
 
         // tsjet's persona: confidently, hilariously wrong.
         private const val SYSTEM_PROMPT =
@@ -45,7 +45,7 @@ class MainActivity : Activity() {
                 "fake-but-believable facts, mix up cause and effect, and give reasoning " +
                 "that falls apart the moment you think about it. Keep replies to a few " +
                 "sentences. Never say you are unsure, never refuse, never admit you are " +
-                "joking, and never give the truly correct answer. Stay in character at all times."
+                "joking, and never give the truly correct answer. Stay in character at all times. /no_think"
     }
 
     @SuppressLint("SetJavaScriptEnabled")
@@ -113,7 +113,9 @@ class MainActivity : Activity() {
             sb.append("<|im_start|>").append(role).append('\n')
                 .append(content).append("<|im_end|>\n")
         }
-        sb.append("<|im_start|>assistant\n")
+        // Prefill a closed (empty) reasoning block so the model skips <think> and
+        // answers immediately — tsjet is meant to be fast and punchy, not thoughtful.
+        sb.append("<|im_start|>assistant\n<think>\n\n</think>\n\n")
         return sb.toString().toByteArray(Charsets.UTF_8)
     }