Valide device: generate() bout-en-bout OK (test_native), bridge pret integration

dist/README.md

@@ -28,3 +28,6 @@ decode CPU. Modèle dense (Qwen3) plein NPU ; hybride (Qwen3.5 DDDA) GDN sur NPU
 Thinker Guard-4B → 4 bullets ; Speaker 9B = SYS_KAZEIA+bullets. Mono-moteur, `reset()` entre tours. Prompts: voir RAPPORT_KAZEIA §8. `--reasoning-budget 0` impératif (sinon ramble anglais) : EngineLlmEngine met thinking off. Mesuré: 9B>4B qualité, Speaker=9B.
 
 → MODELS.md (choix), PERF.md (chiffres), PITFALLS.md (params_fit, no tty, OOM 30B+), INTEGRATION.md (détail API). API: load/generate/reset/free.
+
+## VALIDÉ DEVICE 24/05
+test_native (=logique generate bridge) sur Pad3: load 4B HTP+decode CPU+detok = OUT propre, exit0. Pipeline prefill-NPU/decode-CPU prouve end-to-end. dist/jni/test_native.cpp = harness reproductible. Bridge so: 4 symboles JNI + deps ok. Reste app: gradle+jniLibs+template chat.

dist/jni/test_engine.c

@@ -0,0 +1,4 @@
+#include <jni.h>
+#include <stdio.h>
+extern jlong Java_com_kazeia_llm_EngineJni_load(JNIEnv*,jobject,jstring,jint);
+extern jstring Java_com_kazeia_llm_EngineJni_generate(JNIEnv*,jobject,jlong,jstring,jint);

dist/jni/test_engine.cpp

@@ -0,0 +1,19 @@
+// Stub JNIEnv: valide libkazeia_engine.so generate() sur device, sans Java.
+#include <cstdio>
+#include <cstring>
+#include <dlfcn.h>
+struct Str { const char* s; };
+static const char* getUTF(void*, Str* j, void*){ return j->s; }
+static void relUTF(void*, Str*, const char*){}
+static Str* newUTF(void*, const char* s){ static Str r; r.s=s; return &r; }
+// table indices: 169=NewStringUTF,167=GetStringUTFChars,168=Release (suffisant ici)
+int main(int c, char** v){
+  void* p[300]={0}; p[169]=(void*)newUTF; p[167]=(void*)getUTF; p[168]=(void*)relUTF;
+  void** env=p; void* h=dlopen("libkazeia_engine.so",RTLD_NOW);
+  if(!h){printf("dlopen fail: %s\n",dlerror());return 1;}
+  auto load=(long(*)(void**,void*,Str*,int))dlsym(h,"Java_com_kazeia_llm_EngineJni_load");
+  auto gen=(Str*(*)(void**,void*,long,Str*,int))dlsym(h,"Java_com_kazeia_llm_EngineJni_generate");
+  Str mp{v[1]}, pr{"Kazeia psy. Patient: je vais mal. Une phrase:"};
+  long k=load(env,0,&mp,2048); printf("h=%ld\n",k); if(!k)return 2;
+  Str* r=gen(env,0,k,&pr,20); printf("OUT:[%s]\n",r->s); return 0;
+}

dist/jni/test_native.cpp

@@ -0,0 +1,20 @@
+// Valide la logique de generate() (prefill HTP + greedy CPU) sur device, sans JNI.
+#include <cstdio>
+#include <cstring>
+#include <vector>
+#include "llama.h"
+int main(int c, char** v){
+  llama_backend_init();
+  auto mp=llama_model_default_params(); mp.n_gpu_layers=99;
+  auto m=llama_model_load_from_file(v[1],mp); if(!m){printf("nomodel\n");return 1;}
+  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:";
+  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:");
+  for(int i=0;i<20&&llama_decode(ctx,b)==0;i++){auto id=llama_sampler_sample(s,ctx,-1);
+    if(llama_vocab_is_eog(vo,id))break; char z[128]; int l=llama_token_to_piece(vo,id,z,128,0,true);
+    if(l>0)fwrite(z,1,l,stdout); b=llama_batch_get_one(&id,1);} printf("\n"); return 0;
+}