agno load path skips the index/side-car consistency check used by the other integrations

[Aravindargutus]

Summary

TurboQuantVectorDb._load_from reconstructs its in-memory maps from docstore.json
without calling check_persisted_handles(...). The other three integrations
(langchain, haystack, llama_index) all run that check after loading. As a result,
if index.tvim and docstore.json get out of sync (partial copy, stale backup, manual
edit), agno surfaces an opaque KeyError deep inside search instead of the clean
ValueError at load time that _persist.py was written to provide.

Where

• turbovec-python/python/turbovec/agno.py::_load_from — rebuilds _u64_to_doc,
  _str_to_u64, _content_hashes, _name_to_ids, but never validates the loaded handle
  set against the index.
• Compare with langchain.py::load, haystack.py::load_from_disk,
  llama_index.py::from_persist_path, which all call
  check_persisted_handles(index, ..., what=...).

Impact

Low severity but a real robustness/consistency gap: a desynced agno store fails late and
unclearly (KeyError during query) instead of early and explicitly (ValueError on load),
which is the exact failure mode _persist.py exists to prevent.

Suggested fix

Add check_persisted_handles(self._index, self._u64_to_doc.keys(), what="document") at the
end of _load_from, matching the other integrations.

Environment

• turbovec 0.8.0 (Python), x86_64 Linux, built from source.

[RyanCodrai]

Thanks, good issue, will take a look soon!