fix: normalize message roles for local model backends (llama.cpp, Ollama)

Local backends using strict chat templates (e.g. Mistral 3) rejected
Flynn's Anthropic-style tool_use/tool_result content blocks, causing
'roles must alternate' errors. Added getMessageTextWithTools() and
normalizeMessagesForLocal() to serialize structured blocks to plain
text, drop empty messages, and merge consecutive same-role messages.
Also fixed compaction to ensure kept messages start with user role.

src/context/compaction.ts

@@ -53,6 +53,12 @@ export async function compactHistory(opts: {
   const toCompact = messages.slice(0, -keepCount);
   const toKeep = messages.slice(-keepCount);
 
+  // Ensure toKeep starts with a user message to avoid assistant→assistant
+  // after the compaction summary (which has role 'assistant').
+  while (toKeep.length > 0 && toKeep[0].role === 'assistant') {
+    toCompact.push(toKeep.shift()!);
+  }
+
   const formattedConversation = toCompact.map((msg) => `${msg.role}: ${getMessageText(msg)}`).join('\n\n');
 
   const tier = orchestrator.getDelegationTier('compaction');