Epic: Two-Phase Read Pushdown (predicate/projection/limit/order-by)

[ethe]

Goal

Implement RFC 0010’s two-phase read path so scans push predicates, projection, limit, and PK ordering down to memtables and Parquet SSTs. Deliver a reusable ScanPlan with pruning/RowSets and an execution stream that filters before projection, prunes row-groups/pages, enforces limits early, and preserves PK order.

Current state to build on:

• ScanPlan/projection_with_predicate exist; MergeStream/PackageStream already do PK-ordered MVCC merge + limit; SstableScan handles MVCC/delete sidecar.
• Missing: RowSet abstraction, SST-level pruning (sst_entries includes all), row-group/page pruning, Parquet RowFilter pushdown, bloom filters.

Acceptance criteria:

• Planner builds a RowSet per source, extends scan schema with predicate columns, and filters SST/memtable inputs by key bounds and commit_ts stats.
• Parquet scans fetch metadata async, prune row-groups/pages by min/max (and read_ts), and push a predicate-derived RowFilter; only projected columns are read.
• Execution preserves PK-ascending order through pruning, applies residuals (if any), and enforces limit early (stop once satisfied).
• Missing columns in predicates error cleanly; NULL semantics match RFC.
• Tests cover SST/memtable pruning, row-group/page skip, projection+predicate, MVCC correctness, early limit, PK order contract, and missing-column errors.
• Bloom filters are either implemented as a stretch (write+read) or explicitly out of scope.

Stories:

1. Planner & pruning foundation
2. Parquet pushdown execution
3. Validation & tests
4. Stretch: Bloom filters

[linear]

DEV-509