Bridge v3 utilisable: thinking-off ChatML, sys+usr, cap; static intact

dist/README.md

@@ -34,3 +34,6 @@ test_native (=logique generate bridge) sur Pad3: load 4B HTP+decode CPU+detok =
 
 ## v2 STATIC (collision libllama TTS resolue)
 TTS Talker/CP linke vs son libllama.so -> conflit ABI. FIX: libkazeia_engine.so STATIC (llama+ggml+ggml-cpu+hexagon en .a, --whole-archive hexagon). NEEDED= libm/log/dl/c only, 42MB. Coexiste avec libllama TTS. Drop: libkazeia_engine.so + libggml-htp-v79.so. Valide device 4B HTP. Build: bstatic BUILD_SHARED_LIBS=OFF.
+
+## v3 utilisable: thinking-off bridge
+generate(sys,usr,max): wrap ChatML + <think></think> vide = stop reasoning, decode 4B 15tok/s ~5s/tour. test 4B: "Je suis desole..." 40s(load)+gen. signature 2-arg. cap maxTok 64. dist=libkazeia_engine.so static + htp-v79. #277 fix livre.

dist/jni/EngineLlmEngine.kt

@@ -2,13 +2,13 @@ package com.kazeia.llm
 // Remplace ExecuTorchLlmEngine. Mono-moteur Speaker+Thinker (budget thinking off).
 class EngineJni {
     external fun load(ggufPath: String, nCtx: Int): Long
-    external fun generate(h: Long, prompt: String, maxTok: Int): String
+    external fun generate(h: Long, sys: String, usr: String, maxTok: Int): String
     external fun reset(h: Long); external fun free(h: Long)
     companion object { init { System.loadLibrary("kazeia_engine") } }
 }
 class EngineLlmEngine(model: String, ctx: Int = 2048) {
     private val jni = EngineJni(); private val h = jni.load(model, ctx)
     // Thinker: bullets; Speaker: SYS_KAZEIA + bullets. reset() entre tours, ou cache persistant.
-    fun generate(p: String, max: Int = 64) = jni.generate(h, p, max)
+    fun generate(sys: String, usr: String, max: Int = 64) = jni.generate(h, sys, usr, max)
     fun reset() = jni.reset(h); fun release() = jni.free(h)
 }

dist/jni/kazeia_engine_jni.cpp

@@ -23,11 +23,15 @@ Java_com_kazeia_llm_EngineJni_load(JNIEnv* e, jobject, jstring path, jint nctx)
 }
 
 extern "C" JNIEXPORT jstring JNICALL
-Java_com_kazeia_llm_EngineJni_generate(JNIEnv* e, jobject, jlong h, jstring prompt, jint maxTok) {
-    auto* k = (KEngine*)h; const char* pr = e->GetStringUTFChars(prompt, 0);
-    int n = -llama_tokenize(k->v, pr, strlen(pr), nullptr, 0, true, true);
-    std::vector<llama_token> t(n); llama_tokenize(k->v, pr, strlen(pr), t.data(), n, true, true);
-    e->ReleaseStringUTFChars(prompt, pr);
+Java_com_kazeia_llm_EngineJni_generate(JNIEnv* e, jobject, jlong h, jstring sys, jstring usr, jint maxTok) {
+    auto* k = (KEngine*)h;
+    const char* sp = e->GetStringUTFChars(sys, 0); const char* up = e->GetStringUTFChars(usr, 0);
+    // ChatML Qwen3.5 + <think></think> vide = thinking OFF deterministe (sinon boucle infinie)
+    std::string p = "<|im_start|>system\n"; p+=sp; p+="<|im_end|>\n<|im_start|>user\n"; p+=up;
+    p += "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
+    e->ReleaseStringUTFChars(sys, sp); e->ReleaseStringUTFChars(usr, up);
+    int n = -llama_tokenize(k->v, p.c_str(), p.size(), nullptr, 0, true, true);
+    std::vector<llama_token> t(n); llama_tokenize(k->v, p.c_str(), p.size(), t.data(), n, true, true);
     llama_batch b = llama_batch_get_one(t.data(), n);            // prefill batch -> HTP
     std::string out; char buf[256];
     for (int i = 0; i < maxTok && llama_decode(k->c, b) == 0; ++i) {

dist/jni/test_native.cpp

@@ -10,7 +10,7 @@ int main(int c, char** v){
   auto cp=llama_context_default_params(); cp.n_ctx=2048; cp.n_threads=8; cp.n_batch=512;
   auto ctx=llama_init_from_model(m,cp); auto vo=llama_model_get_vocab(m);
   auto s=llama_sampler_init_greedy();
-  const char* pr="Kazeia psy. Patient: je vais mal. Une phrase:";
+  const char* pr="<|im_start|>system\nKazeia, psy bref FR.<|im_end|>\n<|im_start|>user\nje vais mal<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
   int n=-llama_tokenize(vo,pr,strlen(pr),0,0,true,true); std::vector<llama_token>t(n);
   llama_tokenize(vo,pr,strlen(pr),t.data(),n,true,true);
   auto b=llama_batch_get_one(t.data(),n); printf("OUT:");