fix(render-plan): rewrite raw CTE-name qualifiers to FROM/JOIN alias

[genezhang]

Summary

Unlocks 8 LDBC queries on Spark/Delta (Category C from the sweep): bi-1, bi-2, bi-5, complex-1, complex-4, complex-7, complex-9, complex-11.

Sweep result: 15p / 21xf / 5s → 23p / 13xf / 5s.

Problem

After a WITH→CTE barrier, downstream SELECT items could be emitted qualified by the CTE table name (with_friend_cte_1.p6_friend_id) even when the same CTE was JOINed under a user-facing alias (INNER JOIN with_friend_cte_1 AS friend). ClickHouse silently accepts either qualifier; Spark/Delta resolves only against the alias actually bound in FROM/JOIN, producing UNRESOLVED_COLUMN.

Two render-plan construction paths (rewrite_with_aliases_to_cte, rewrite_table_aliases_to_cte) build PropertyAccess { table_alias: cte_name, ... } directly. Both leaked CH-only resolution behaviour into the emitted SQL.

Fix

Extend fix_orphan_table_aliases_impl (src/render_plan/variable_scope.rs) so its alias_replacements map also includes cte_name → from_alias entries for any CTE referenced from FROM/JOIN. The existing rewrite_expr_table_aliases walker then normalizes any stray CTE-name qualifier in SELECT/WHERE/GROUP BY/ORDER BY/HAVING/JOIN-ON to the bound alias.

• Skips entries where cte_name == from_alias (no-op)
• Uses entry().or_insert_with so prior orphan-alias mappings win on collision

Stragglers

Two queries that were tagged [C] in PR #347 remain failing for distinct reasons (now retagged):

• bi-14 — same alias person1 rebound across 5 chained CTEs. person1.score in the 5th CTE doesn't resolve against the 4th CTE's schema.
• complex-3 — t5.CountryId emitted for a Place_isPartOf_Place rel (no such column — Place→Place edge uses PlaceId).

Test plan

• cargo fmt --all && cargo clippy --all-targets --features databricks clean
• cargo test --lib — 1370 passed
• CLICKGRAPH_SPARK_TESTS=1 pytest tests/spark_smoke/test_ldbc_sweep.py — 23p / 13xf / 5s (was 15/21/5); no regressions on the original 15 passing
• pytest --runxfail confirms the 8 newly-passing queries return correct results, not just no-error

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes a Spark/Delta SQL compatibility issue in render-plan aliasing: after a WITH→CTE barrier, expressions could be qualified by the raw CTE table name even when the CTE was bound under a different FROM/JOIN alias, which Spark/Delta does not resolve (but ClickHouse does). The change extends the final alias-rewrite pass so stray with_*_cte_N.col qualifiers are normalized to the actual bound FROM/JOIN alias, unlocking additional LDBC queries in the Spark smoke sweep.

Changes:

• Extend fix_orphan_table_aliases_impl to also rewrite raw CTE-name qualifiers (cte_name → from/join alias) in downstream expressions.
• Update the Spark LDBC sweep xfail list to drop the 8 queries now passing and retag remaining Category C failures with more specific reasons.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/render_plan/variable_scope.rs	Adds CTE-name→FROM/JOIN-alias entries to the alias replacement map so expression qualifiers match the bound alias on Spark/Delta.
tests/spark_smoke/test_ldbc_sweep.py	Updates EXPECTED_FAILURES to reflect the newly passing queries and reclassifies remaining C-resolution failures.