Fix silent row loss when LOAD FROM/UNWIND feeds a MATCH primary key predicate

[adsharma]

Summary

Fixes #861

LOAD FROM <source> followed directly by MATCH (n:Label {pk: col}) — where col is a scan column — matched zero rows silently on 0.19.0–0.20.0. No error was raised, the query reported success, and any SET/MERGE hanging off the MATCH never executed:

LOAD FROM 'updates.csv' (header=true)
MATCH (p:Products {handle: handle})
SET p.description = description
RETURN p.handle   -- 0 rows; p.description unchanged

Root cause

The planner legitimately rewrites a single-node MATCH with a primary-key equality predicate into a QUERY_PRIMARY_KEY_LOOKUP whose key is evaluated on top of the outer (scan) pipeline. However, ProjectionPushDownOptimizer never visited LogicalQueryPrimaryKeyLookup, so its key expression was never collected as in-use. visitTableFunctionCall then marked the referenced scan column as skippable (columnSkips) and the physical scan stopped writing that vector. The key evaluator subsequently read an unwritten vector, the key evaluated to NULL for every row, lookupPK failed, and the lookup returned an empty result — silently.

This also explains the report's observations:

• LOAD FROM df MATCH (p:Products {handle: handle}) SET ... failed silently (the key column was pruned from the scan).
• An explicit WITH handle AS h between the scan and the MATCH worked: the intermediate PROJECTION starts a fresh push-down pass that keeps the column alive.
• Matching on a non-primary-key property ({description: description}) worked: it becomes a regular filter, which collects its predicate.
• MATCH (p:Products {handle: 'a'}) (literal key) worked: no outer dependency, no pruning.

Fix

Override visitQueryPrimaryKeyLookup in ProjectionPushDownOptimizer to collect the lookup's key and node-ID expressions as in-use, so the scan column they reference is kept alive. This works for both the raw-variable case (the issue's {pk: col}) and arbitrary key expressions (e.g. the implicit CAST inserted for a SERIAL primary key), since collectExpressionsInUse recurses into children.

Testing

• Added regression cases to test/test_files/load_from/load_from.test:
  • LOAD FROM ... MATCH (p:person {ID: id}) RETURN COUNT(*) → 5 (property-map form)
  • LOAD FROM ... MATCH (p:person) WHERE p.ID = id RETURN COUNT(*) → 5 (equivalent WHERE form)
  • LOAD FROM ... MATCH (p:person {ID: id}) SET p.fName = fName RETURN p.fName ORDER BY id → 5 rows (verifies SET executes through the lookup)
• e2e_test full run: 1960 passed; the only 2 failures (copy_to_csv.CopyToInvalidCase, dictionary_bug~...AnonymousParquetDeleteReload) are pre-existing environment issues (missing dataset/empty fixtures; JSON export extension not built) and fail identically without this change.
• optimizer_test, planner_tests, binder_test, api_test all pass.