reindex --search cannot repair DB-only search index divergence #1226

Problem

basic-memory reindex --search cannot repair database-only divergence: it selects files to
index purely from file-content checksum changes, so when the observation/relation/
search_index tables disagree with an unchanged markdown file, the incremental pass selects
zero files and reports success while search keeps serving stale rows.

Field report on #1214 (SQLite, concurrent MCP clients): after trimming duplicate
observation rows, search_index still held 3886 rows of type='observation' against 2146
observation rows. reindex --search reported Indexing files... 0/0 files and changed
nothing; until reindex --full --search ran, search answered with content no longer in the
store.

Root cause

select_project_index_target_files (src/basic_memory/indexing/project_index_coordinator.py:229-242)
builds its target set from change_report.new_files | modified_files, which
change_detector.py derives purely from file checksums (load_indexed_file_checksums).
DB-only divergence leaves every file in unchanged_files, so nothing reaches
search_service.index_entity_data — the only code path that deletes and rewrites an
entity's observation search rows. _purge_stale_search_rows only handles rows for deleted
entities, and only runs in the --embeddings path.

Proposed direction

Add a DB↔FTS reconciliation sweep to the search reindex path (extend
_purge_stale_search_rows or a sibling): purge search_index rows whose id no longer
exists in the owning table (type='observation' → observation.id, type='relation' →
relation.id), and optionally detect entities whose row counts disagree and re-enqueue just
those entities, so --search without --full converges without a full rebuild.

The one-time damage from the observation duplication bug is handled by the #1214 migration's
FTS orphan purge; this issue is about making incremental reindex --search able to repair
such divergence in general.

Acceptance criteria

• reindex --search (no --full) removes search_index rows of type='observation'/type='relation' whose backing row no longer exists.
• A seeded DB↔FTS divergence on an unchanged file converges via reindex --search alone (regression test).
• No full-file re-parse is required for the purge path; --full behavior unchanged.

Refs: #1214 (field report), #1213.