fix(quick-1): bail on NULL KV cache and improve prefill error diagnostics

- engine.rs: new_session() now returns Err when llama_get_memory returns NULL
  instead of silently continuing, preventing decode corruption on 2nd request
- session.rs: decode_prefill error includes token count for easier diagnosis

crates/inference-core/src/engine.rs

@@ -263,10 +263,6 @@ impl InferenceEngine {
         info!("Running engine warmup (dummy decode)…");
 
         unsafe {
-            // Mark context as warmup so llama.cpp skips perf counters
-            if let Ok(f) = std::panic::catch_unwind(|| {}) {
-                let _ = f;
-            }
             (self.lib.set_warmup)(self.ctx.ptr, true);
             // set_warmup may not exist in all builds; call decode with 1 token
             let dummy_token: llama_sys::llama_token = 1; // BOS-ish token
@@ -279,7 +275,11 @@ impl InferenceEngine {
             }
 
             let memory = (self.lib.get_memory)(self.ctx.ptr);
-            (self.lib.memory_clear)(memory, false);
+            if memory.is_null() {
+                warn!("llama_get_memory returned NULL during warmup — KV cache not cleared");
+            } else {
+                (self.lib.memory_clear)(memory, false);
+            }
         }
 
         info!("Warmup complete");
@@ -295,6 +295,21 @@ impl InferenceEngine {
         prompt: &str,
         sampler_cfg: SamplerConfig,
     ) -> Result<GenerationSession> {
+        // Clear KV cache from any previous session so the new prefill starts at position 0.
+        // `true` forces a full clear including M-RoPE position metadata; `false` only clears
+        // cell occupancy flags and leaves position data intact, which causes the
+        // "inconsistent sequence positions" error on models that use M-RoPE (e.g. Qwen).
+        unsafe {
+            let memory = (self.lib.get_memory)(self.ctx.ptr);
+            if memory.is_null() {
+                tracing::error!("llama_get_memory returned NULL — cannot clear KV cache, refusing to create session to avoid decode corruption");
+                anyhow::bail!("KV cache clear failed: llama_get_memory returned NULL");
+            } else {
+                (self.lib.memory_clear)(memory, true);
+                tracing::info!("engine.kv_cache_cleared");
+            }
+        }
+
         let vocab = unsafe { (self.lib.model_get_vocab)(self._model.ptr) };
         GenerationSession::with_vocab(
             Arc::clone(&self.lib),