feat(sql): support streaming IN subqueries with mark joins

[ksushant6566]

Adds the remaining dynamic IN (SELECT ...) case from #1659 through a two-input mark join in the streaming DAG. An inner insert can admit already-seen outer rows; deleting the final matching inner value retracts them. Inner duplicates preserve membership without multiplying outer rows.

SELECT * FROM orders
WHERE customer_id IN (
  SELECT customer_id FROM customers WHERE active = 1
);

The join retains outer-row multiplicity and reference-counts inner values. It applies complete update/batch deltas before emitting nullable membership changes. Same-type keys use a hash index; cross-type values use Dozer's SQL equality conversions, rejecting incomparable non-null operands explicitly. The normal selection stage evaluates compound predicates and removes internal mark fields before projection or aggregation, so SELECT * retains its original schema.

The planner handles multiple and nested IN/NOT IN predicates, scalar left operands, CTEs and parenthesized UNION branches. Supporting nullable membership also required correcting AND/OR three-valued evaluation and nullable metadata. Qualified column resolution now checks the qualifier even when only one field name matches, so a correlated reference cannot silently bind to an inner column. Query-local CTE names are restored at the query boundary instead of leaking into sibling subqueries.

Scope: uncorrelated, single-column subqueries. TTL-bearing inputs are rejected before any state mutation because existing TTL index eviction emits no delete events; silently expiring membership would leave downstream rows stale. Existing ORDER BY/LIMIT restrictions remain. The prior static-list IN implementation is retained.

Demo

Watch the approximately 34-second MP4, or download the original terminal recording.

This executes the actual parser, schema validation, generated DAG and processors with in-memory source events. It shows late inner matches, duplicate suppression, atomic key updates and retractions. Run cargo test -p dozer-sql --no-default-features in_subquery_demo -- --nocapture to reproduce it. Details are in dozer-tests/in_subquery/README.md.

Validation

• On clean upstream 8b0b4cc, the new parser/DAG regression fails with UnsupportedExpression(InSubquery ...).
• cargo test --locked -p dozer-sql -p dozer-sql-expression --no-default-features: 238 SQL tests and 27 expression tests pass; one pre-existing SQL test remains ignored. Doc tests pass.
• cargo check --locked -p dozer-cli --no-default-features passes with Rust 1.78, including the full CLI dependency graph.
• 24 new SQL-to-DAG cases cover both arrival orders, all outer WHERE update transitions, inner key updates, identical updates, duplicate/NULL batches, empty-set and literal-NULL semantics, compound negation, numeric/string conversion, incompatible comparisons, nested/parenthesized queries, CTE shadowing, shared sources and TTL rejection without state mutation. A deterministic oracle checks 360 successive stream changes.
• Independent agent review supplied additional regressions, reproduced the NULL/qualifier/scope failures and verified their fixes.
• cargo clippy --locked -p dozer-sql -p dozer-sql-expression --no-default-features --all-targets succeeds with warnings, including the existing large error types. Rustfmt and git diff --check pass.

AI assistance: implemented and tested with Codex.

Fixes #1659
/claim #1659

[CLAassistant]

All committers have signed the CLA.