fix: handle multiple root-level XML elements from non-conformant LLMs

[octo-patch]

Fixes #407

Problem

Some LLMs (e.g. llama3.2 via Ollama) do not follow the expected XML output format precisely. Instead of wrapping all extracted memories in a single <item> root element, they produce one <item> element per memory with natural-language numbering in between:

1. <item>
    <memory>...</memory>
</item>

2. <item>
    <memory>...</memory>
</item>

_find_xml_boundaries extracts the slice from the first <item> to the last </item>, which contains multiple top-level tags separated by text (2. , 3. , …). ET.fromstring then raises ParseError: junk after document element because the first root element closes before the end of the string.

Solution

Two minimal changes to _parse_memory_type_response_xml:

1. Wrap-and-retry: catch ParseError from the first ET.fromstring call, wrap the content in a synthetic <_root_> element, and retry. This makes the multi-<item> layout valid XML without affecting well-formed single-root responses.

2. findall → iter: switch from root.findall("memory") to root.iter("memory") so that <memory> nodes are found regardless of depth—direct children of the root (normal case) or grandchildren via <item> wrappers (wrapped multi-root case).

Testing

Manually validated against the XML snippets from the issue report. The retry path successfully parses multiple top-level <item> elements and returns all contained <memory> items.