fix(qwen35): early SSM probe fixes DeltaNet layer detection — Qwen3.5-4B now coherent

The fused QKV guard (delta_a_log check) ran BEFORE delta_a_log was
set in the per-layer loop, so DeltaNet layers still got gguf_w_qkv
assigned. Fixed by probing blk.N.ssm_a BEFORE the fused QKV check.

Result: n_attn_layers correctly reports 8 (not 32) for Qwen3.5-4B.
DeltaNet layers dispatch to deltanet_forward, full attention layers
to self_attn_forward with partial RoPE + NeoX rotation.

Validated:
  Qwen3.5-4B: 'The capital of France is **Paris**.'
  Qwen3-4B:   coherent (4.5 tok/s)
  Phi-3.5:    coherent (1.9 tok/s)
  35/35 unit tests pass

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

quant.h

@@ -11851,20 +11851,20 @@ tq_model_t* tq_load_gguf(const char* path) {
             }
         }
 
-        /* Phi-3 fused QKV detection.
-         *
-         * Phi-3 ships `blk.N.attn_qkv.weight` with shape [hidden, 3*hidden]
-         * instead of three separate `attn_q/k/v.weight` tensors. We store
-         * the fused pointer in `gguf_w_qkv` and the forward path dispatches
-         * one matmul + split. The layer is marked as an attention layer
-         * via the same `is_attn_layer` flag the standard path uses, so
-         * the rest of the loader and tq_forward treat it normally. */
+        /* Early DeltaNet probe: check if this layer has SSM weights BEFORE
+         * the fused QKV detection. DeltaNet layers also have attn_qkv.weight
+         * (for conv1d input), and we must NOT treat it as a Phi-3 fused QKV. */
+        int layer_is_deltanet = 0;
+        {
+            char ssm_probe[128];
+            snprintf(ssm_probe, sizeof(ssm_probe), "blk.%d.ssm_a", l);
+            if (find_gguf_tensor(gguf, ssm_probe)) layer_is_deltanet = 1;
+        }
+
+        /* Phi-3 fused QKV detection (skip for DeltaNet layers). */
         snprintf(tname, sizeof(tname), "blk.%d.attn_qkv.weight", l);
         const tq_gguf_tensor_t* wqkv_t = find_gguf_tensor(gguf, tname);
-        if (wqkv_t && !layer->delta_a_log) {
-            /* Phi-3 fused QKV (NOT DeltaNet). DeltaNet layers also have
-             * attn_qkv.weight but it's the conv1d input, not a fused
-             * attention projection. The delta_a_log check distinguishes. */
+        if (wqkv_t && !layer_is_deltanet) {
             layer->gguf_w_qkv = wqkv_t->data;
             layer->gguf_w_qkv_type = wqkv_t->type;
             c->has_fused_qkv = 1;