fix: keep SQLite fuzzy-term index memory in sync with the database

[Kevin Turcios (KRRT7)]

Summary

SqliteRelatedTermsFuzzy keeps a VectorBase, a term list (vector position → term), and an added-terms set in memory alongside the RelatedTermsFuzzy table. Three paths let the in-memory state drift from the table:

• clear() only deleted the rows. Lookups still returned cleared terms, and re-adding a cleared term was skipped as "already added", so the table stayed empty (size 0, get_terms() == []).
• deserialize() of empty data returned before clearing the VectorBase. Terms added afterwards were appended after the stale vectors, so their positions no longer matched the term list and lookups could not find them.
• Opening an existing database loaded stored terms into the term list but not the added-terms set, so re-adding a stored term duplicated it: lookup_term("plum") returned plum twice and serialize() listed it twice.

The fix resets all three in-memory structures in clear() and deserialize(), and populates the added-terms set when loading.

Validation

• New tests for each path: clear then re-add, empty deserialize then add, and reopen then re-add. Each fails without the fix.
• make format check test

[Bernhard Merkle (bmerkle)]

LGTM — verified locally.

• All three drift paths reproduce: each new test fails against main's related_terms_index.py and passes with the fix. Full make is green (814 passed).
• Checked the rest of the class for the same class of bug: remove_term raises NotImplementedError, and the add_terms_with_embeddings and non-empty deserialize paths already keep _vector_base / _terms_list / _added_terms in step with the table.

Non-blocking follow-up, same family: duplicates within one batch aren't filtered. add_terms(["fig", "fig", "date"]) writes 2 rows (thanks to INSERT OR REPLACE), but _terms_list/VectorBase get 3 entries, so serialize() returns ['fig', 'fig', 'date'] and lookup_term("fig") returns fig twice. Production callers pass unique terms today, so it's latent; de-duplicating new_terms in add_terms_with_embeddings (e.g. via dict.fromkeys) would close it.