neo4j: Data_Bank is set once on create, never reset by a re-ingest

[Evanfeenstra]

NodeQueryBuilder::build() put Data_Bank = name into $properties and ran ON MATCH SET node += $properties, so an incremental sync of a file reset every one of its nodes' search text to the bare name while text_embeddings and _search_fields_used survived — a vector that no longer matched its text, and jarvis's data_bank_full_index back to name-only (Problem 5 in stakwork/strut plans/semantic-search.md; built by stakwork/strut#140).

Data_Bank now stays out of $properties and is seeded from the map's name under ON CREATE alone (node.Data_Bank = $properties.name — no new parameter; the binder already passes the whole map), like date_added_to_graph: the same set-once rule, for the same reason. A streamed node is searchable at once as before; a later writer — strut's embedding pass, jarvis — composes the richer text from the type's Schema index fields, and a re-ingest of an unchanged function touches nothing it reads. An unnamed node seeds no search text, as before. build_stream() was already ON CREATE only and is untouched.

Without this strut is still correct (its pass sees the reset text as stale and re-embeds the node), but a full re-ingest would re-embed every code node of the repository each time.

Test: build_sets_data_bank_only_on_create_and_keeps_it_out_of_properties, beside the existing set-once tests (cargo test -p ast --lib --features neo4j build_sets). cargo fmt --check on the file reports a pre-existing diff in read_probe_date_added, not touched here.

Also bumps mcp's strut pin to a50eb64, the merge of stakwork/strut#140 (the only changes between the previous pin and it are that PR and its plan documents, #138 and #139). So this is the pin that puts the embedding pass, the Schema index reconcile and the typed-search routing on every swarm at its next deploy — the swarm38 validation in the plan starts after it.