feat: add Qwen3.6 models and refresh model database with discovery

Add Qwen3.6-27B (dense, 27.8B) and Qwen3.6-35B-A3B (MoE, 36B/3B active)
to the curated model list with GGUF sources from unsloth and ggml-org.
Add hybrid attention layout inference for Qwen3.5/3.6 architectures.
Full re-scrape with --discover -n 2000 brings total to 962 models with
422 GGUF-sourced entries.

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

llmfit-core/src/models.rs

@@ -845,6 +845,23 @@ pub fn infer_attention_layout_from_name(name: &str) -> Option<AttentionLayout> {
         });
     }
 
+    // Qwen3.5 / Qwen3.6 hybrid models use 1 full attention per 4 layers.
+    // The dense 27B variants have 64 layers → 16 full + 48 linear.
+    // The MoE A3B variants have 40 layers → 10 full + 30 linear.
+    if lower.contains("qwen3.5-") || lower.contains("qwen3.6-") {
+        if lower.contains("-a3b") || lower.contains("-a10b") || lower.contains("-a17b") {
+            return Some(AttentionLayout {
+                full: 10,
+                linear: 30,
+            });
+        }
+        // Dense variants (27B) use 64 layers with same 1:3 ratio
+        return Some(AttentionLayout {
+            full: 16,
+            linear: 48,
+        });
+    }
+
     // Jamba (Mamba + Transformer hybrid). Jamba 1.5 Mini and Large both
     // use a 1:7 attention to mamba ratio in their 32 layer blocks.
     if lower.contains("jamba") {
@@ -1722,7 +1739,7 @@ mod tests {
             "meta-llama/Llama-3.3-70B-Instruct",
             "Qwen/Qwen2.5-7B-Instruct",
             "Qwen/Qwen2.5-Coder-7B-Instruct",
-            "meta-llama/Meta-Llama-3-8B-Instruct",
+            "meta-llama/Llama-3.1-8B-Instruct",
             "mistralai/Mistral-7B-Instruct-v0.3",
         ];
         for name in &expected_with_gguf {

scripts/scrape_hf_models.py

@@ -94,6 +94,9 @@ TARGET_MODELS = [
     "Qwen/Qwen3.5-2B-Base",
     "Qwen/Qwen3.5-4B-Base",
     "Qwen/Qwen3.5-9B-Base",
+    # Qwen 3.6 (native multimodal + hybrid attention, Apr 2026)
+    "Qwen/Qwen3.6-27B",
+    "Qwen/Qwen3.6-35B-A3B",
     # Microsoft Phi
     "microsoft/phi-3-mini-4k-instruct",
     "microsoft/Phi-3-medium-14b-instruct",
@@ -296,6 +299,7 @@ MOE_ACTIVE_PARAMS = {
     "Qwen/Qwen3.5-35B-A3B": 3_000_000_000,
     "Qwen/Qwen3.5-122B-A10B": 10_000_000_000,
     "Qwen/Qwen3.5-397B-A17B": 17_000_000_000,
+    "Qwen/Qwen3.6-35B-A3B": 3_000_000_000,
     "meta-llama/Llama-4-Scout-17B-16E-Instruct": 17_000_000_000,
     "meta-llama/Llama-4-Maverick-17B-128E-Instruct": 17_000_000_000,
     "xai-org/grok-1": 86_000_000_000,