add_system_prompt() modified the message list even when the memory
lookup returned nothing (memories == "").
- If an existing system prompt was present, " \n " (whitespace) was
appended to its content.
- If no system prompt existed, a new {"role": "system", "content": ""}
message was prepended.
Both cases pollute the conversation context with empty/whitespace tokens
that waste prompt budget and can confuse models sensitive to blank
system messages.
Fix: return messages unchanged when memories is falsy.
Adds a regression test covering both cases (with and without existing
system prompt).
