At 327 lines it interleaved reading the tables, deriving the lookups every
check needs, and the bodies of ten checks. Six checks were already
functions; the rest were inline, so none of them could be read or tested
without the others around them.
Each one is now a function taking exactly what it needs:
_check_document_meta_parity, _check_orphaned_chunks, _check_orphaned_items,
_check_documents_without_items, _check_dangling_item_refs,
_check_vector_dimension, _check_unembedded_chunks, _check_picture_data,
_check_settings_row and _check_pending_migrations. run_db_checks reads the
tables, then appends results.
_document_centroids takes the vector reduction. Passing the matrix as a
parameter keeps it a local of run_db_checks, so the del before clustering
still drops the last reference — measured at 6.2 MB allocated to reduce a
102 MB matrix, no second copy. There is no snapshot object: one holding
vectors would keep the largest allocation alive past the del.
The reduction also rebound doc_ids from the document-id set to the centroid
id list halfway through the function. The centroid ids have their own name
now.
No test changes: the 80 doctor tests cover these through run_db_checks and
pass unchanged.
0592206f16