fix: let an unchanged conversation hit the prompt cache

src/common/AutoModel/automodel.cpp

@@ -198,6 +198,13 @@ bool AutoModel::_shared_insert(chat_meta_info_t& meta_info, std::vector<int>& to
         clear_context();
         skip_count = 0;
     }
+    // A fully cached prompt leaves nothing to prefill, and _chunked_insert then
+    // computes zero chunks and returns a default-constructed (empty) logits
+    // buffer straight into sampler->sample(). Re-prefill instead.
+    if (skip_count == tokens.size()) {
+        clear_context();
+        skip_count = 0;
+    }
     tokens.erase(tokens.begin(), tokens.begin() + skip_count);
 
     if (this->total_tokens + tokens.size() >= this->MAX_L){

src/include/prompt_cache.hpp

@@ -173,9 +173,11 @@ class PromptCache {
         // (in order) at the start of the new conversation, allowing rounds
         // produced by other backends (cloud) to be appended without
         // invalidating the locally-built KV cache prefix.
-        const size_t prefix_len = messages.size() - 2;
+        // Bound by the full incoming length, not length-2. Requiring two *new*
+        // messages made a resend of an unchanged conversation miss every time,
+        // which is exactly what a client sends after a timeout.
         const bool can_use_message =
-            message_checksums_.size() <= prefix_len &&
+            message_checksums_.size() <= messages.size() &&
             matched == message_checksums_.size();
         const bool can_use_tools = tool_checksums_ == new_tool_checksums;
         info.tools_matched = can_use_tools;