[MayCXC]

Deleting a DiskANN column's medoid while rows wait in its buffer leaves the graph unsearchable until the buffer is flushed.

diskann_medoid_handle_delete() picks the new medoid with SELECT rowid FROM _vectorsNN WHERE rowid != ?1 LIMIT 1. With buffer_threshold set, _vectorsNN also holds the buffered rows, which have no node yet, and the lowest rowid can be one of them. Every search then starts from a node diskann_search() cannot read: the loop skips it and ends, the graph contributes no rows, and the KNN query returns only what the buffer scan finds.

create virtual table t using vec0(emb float[8] indexed by diskann(neighbor_quantizer=int8, n_neighbors=8, buffer_threshold=10));
-- rows 100 to 109 (the tenth insert flushes them into the graph), then rows 1, 2 and 3 (buffered)
delete from t where rowid = 100;  -- the medoid
-- the new medoid is 1, which has no node
select rowid from t where emb match :vector_of_101 and k = 1;  -- returns a buffered row, not 101

On main, after that delete, none of the 9 remaining nodes is returned by a k = 1 query with its own vector, for each of eight random tables. Here, the new medoid is the first other node of _diskann_nodesNN, so it is always a node of the graph, and all 9 are found in each of them. When no node is left, the medoid is set to empty as before. A failed step of that query is now returned, where it used to set the medoid to empty as well.

test_diskann_delete_medoid_with_rows_in_buffer in test-diskann.py fails on main and passes here; the Python suite passes (240 passed, 66 skipped), test-diskann.py passes against an AddressSanitizer and UndefinedBehaviorSanitizer build, and the build's warnings are main's. The test uses the int8 quantizer: under the binary quantizer a query can also lose its own row to one whose quantized distance ties it, which #345 fixes.

It merges cleanly with every open PR that touches the DiskANN code, #344 and #345 included.

Origin. Found in the DiskANN port to photostructure/sqlite-vec, photostructure#11 (cd53413), which lists it among the port's differences from upstream; the change and its test are written here for asg017's main. No open or closed PR or issue here makes the same change.