dflash: read the confidence-head metadata key instead of only guessing

src/llama-arch.cpp

@@ -340,6 +340,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
 
     { LLM_KV_TARGET_LAYERS,         "%s.target_layers"        },
+    { LLM_KV_CONFIDENCE_HEAD,       "%s.confidence_head"      },
     { LLM_KV_LOG_SNR_CONDITIONING,  "%s.log_snr_conditioning" },
     { LLM_KV_MIN_LOG_SNR,           "%s.min_log_snr"          },
     { LLM_KV_MAX_LOG_SNR,           "%s.max_log_snr"          },

src/llama-arch.h

@@ -386,6 +386,7 @@ enum llm_kv {
     LLM_KV_CLASSIFIER_OUTPUT_LABELS,
 
     LLM_KV_TARGET_LAYERS,
+    LLM_KV_CONFIDENCE_HEAD,
     LLM_KV_LOG_SNR_CONDITIONING,
     LLM_KV_MIN_LOG_SNR,
     LLM_KV_MAX_LOG_SNR,

src/models/dflash.cpp

@@ -118,7 +118,17 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
     //
     // TODO: only Qwen3-style backbones are supported for now; other backbones (e.g. Gemma4)
     //       need their own conversion path and graph tweaks
+// Reject a declared confidence head when its required Markov head is missing.
+    bool kv_confidence_head = false;
+    const bool has_kv_confidence_head = ml->get_key(LLM_KV_CONFIDENCE_HEAD, kv_confidence_head, false);
+
     const struct ggml_tensor * markov_meta = ml->get_tensor_meta("markov_w1.weight");
+
+    if (has_kv_confidence_head && kv_confidence_head && !markov_meta) {
+        throw std::runtime_error("dflash: metadata declares a confidence head, but markov_w1.weight is missing. "
+                                 "The export is incomplete; it would load as plain DFlash and read drafts one row late");
+    }
+
     if (markov_meta) {
         const int64_t dspark_markov_rank = markov_meta->ne[0];