refactor(query_plan): keep from_source and legacy FROM fields in sync (P3 groundwork)

[TimelordUK]

Groundwork for P3 — correlated subqueries (docs/SQL_PARITY.md). No parity behaviour changes here; this makes from_source trustworthy before P3 starts depending on it.

The problem

The executor (query_engine.rs:1217) reads from_source first and only falls back to the deprecated from_table / from_subquery. The parser populates both. Several transformers rewrote only one — so the stale copy is the one that actually executes.

Site	Desync
in_operator_lifter.rs:218	from_table → lifted CTE, from_source → original base table (no lifted column)
expression_lifter.rs:187	same
cte_hoister.rs:61	mutates from_subquery; from_source::DerivedTable keeps a pre-hoist clone
into_clause_remover.rs:44	same

The fix

Adds impl SelectStatement (there was none) with set_from_table, set_from_subquery, map_from_subquery, and routes all four sites through them so the representations can only move together.

map_from_subquery replaces the take-then-restore pattern — that left from_source stale in between and dropped the derived-table alias.

Also deletes column_dependency_lifter.rs: absent from query_plan/mod.rs, referenced nowhere, and no longer compiles against the current AST. Removes 9 deprecated-field sites and 14 SqlExpression match arms from the later P3 audit.

Honest scoping

Neither desync currently produces wrong results. The stale copies happen to be valid statements — pre-hoist from_source still carries its own CTEs, so the executor gets a correct answer by a different route. I tried and failed to build an end-to-end query that exposes either. This pins the invariant rather than fixing live breakage.

Both regression tests parse real SQL instead of hand-building an AST — the existing fixtures set from_source: None, an input the parser never produces, which is exactly why the desync survived. Verified both fail without the fix:

from_source holds a stale pre-hoist subquery with 1 CTE(s)
from_source must follow from_table to the lifted CTE

Verification

• cargo test: 672 passed (+2). One pre-existing failure, history_protection_integration, unrelated — it sets APPDATA to a tempdir but still loads the real roaming history.
• DuckDB parity --check: contract holds, unchanged at 64 AGREE / 1 DIFFER / 7 GAP.
• Python: 497 passed / 18 failed — verified identical on a stashed baseline (missing local data files, offline localhost endpoints).
• Examples: all FORMAL tests pass.
• cargo fmt --check clean; no new clippy warnings in touched files.

Second commit

tests/integration/test_examples.py hardcoded ./target/release/sql-cli, aborting the whole examples suite on Windows before running a case. Mirrors the resolution already in tests/comparison/engines.py. Unrelated to the refactor — split out so it can be dropped independently.

Not addressed (deliberately)

The wider from_source migration. It is blocked on real AST work — TableSource has no table-function variant (recursive_parser.rs:1132) and TableSource::Table cannot carry an alias — which would drag JoinClause in too. Bigger than P3 itself; revisit separately.

🤖 Generated with Claude Code