fix(store): preserve embedding on metadata-only updates

[doctatortot]

Problem

LanceStore::update builds its row via memory_to_batch(memory, vector), and memory_to_batch writes a zero vector when vector is None:

None => vec![0.0; self.vector_dim as usize],

The update handler passes vector = None for any edit that does not change content (tags, state, supersede, …). Combined with merge_insert(...).when_matched_update_all(None), this silently overwrites the row's real embedding with zeros. The memory still exists and reads back fine over the API, but its vector is gone — so it is invisible to vector search, with no error and an HTTP 200.

In practice this wipes embeddings on exactly the memories edited most often (tag/state churn), and the degradation is invisible until you notice recall has quietly rotted.

Fix

When vector is None, reuse the row's existing embedding via get_vector_by_id instead of letting memory_to_batch zero it.

Test

test_metadata_only_update_preserves_vector: create with a vector → metadata-only update (vector = None) → assert the stored vector is unchanged (and non-zero).

[doctatortot]

CI here is red only because of pre-existing rustfmt drift on main (the fmt step fails before build/clippy/test can run). The files changed in this PR are fmt-clean. #27 fixes that drift — once it merges, a rebase onto main turns this green.

[yhyyz]

Merged and deployed — great catch! This was a serious silent data corruption bug introduced by the merge_insert migration. Metadata-only updates were quietly wiping embeddings. Now fixed. 🎉