refactor: slim runtime and bind ORM through planner

[KKould]

Summary

This PR prepares KiteSQL 0.3.0 around a smaller embedded footprint and a cleaner planner/runtime boundary. The main change is not a single executor rewrite: it is a broad cleanup that feature-gates optional frontends, routes ORM queries through the normal binder/planner path, and centralizes execution metadata/codec state so less work is duplicated across storage backends and query shapes.

What Changed

• Feature-gate optional components and dependencies: parser, shell, copy, time, decimal, python, rocksdb, and lmdb; remove the old pgwire server binary and net feature; make kitesql-shell an explicit shell feature target.
• Rework database state around direct metadata caches plus a shared TableArena/PlanArena, so table, view, function, column, and index metadata can be referenced consistently during binding, optimization, and execution.
• Refactor execution setup around ExecutionContext, ExecArena, local execution state, reusable TableCodec, projection scratch buffers, and DDL-apply records. Executor and transaction types are still generic where storage iterators require it, but most nodes now pull shared runtime state from the arena instead of carrying duplicate context.
• Move SQL AST binding behind the parser feature and add binder-facing APIs that build LogicalPlan values directly.
• Rewrite the ORM query/mutation path to use Database::bind / DBTransaction::bind with OrmContext, so typed ORM queries bind into core ScalarExpression and LogicalPlan structures instead of building SQL/parser-shaped intermediate paths.
• Replace bincode/serde-based internal catalog and tuple serialization paths with project-local reference serialization and table-codec implementations, including buffer reuse for tuple decoding/encoding.
• Clean up optimizer and expression execution around evaluator dispatch, column-pruning bookkeeping, parameterized index probes, subquery/set-membership handling, DESCRIBE column references, and allocation reduction.
• Update README, ORM docs, examples, wasm tests, sqllogictest coverage, macro tests, and TPCC benchmark docs for the new APIs and feature layout.

Compatibility Notes

• Bumps the crate version to 0.3.0.
• The kite_sql pgwire server binary and net feature are removed in this PR.
• parser remains enabled by default, but SQL parser support can now be left out of no-default-feature embedded builds.
• orm now enables macros; parser, copy, decimal, time, python, shell, and storage backends are controlled by their own feature flags.
• The documented ORM query-builder style changes from the previous from(...).eq(...).fetch(...) chain to binder-backed bind(|ctx| ...) plans.

Binary Size Comparison

Measured as stripped release binaries on x86_64-unknown-linux-gnu with rustc 1.88.0. The KiteSQL binary is a small ORM/typed API smoke example that creates a table, inserts two rows, fetches them, and drops the table.

Important: the PR measurements intentionally do not enable the parser feature. main still has parser/sqlparser as a non-optional dependency, so main cannot be measured with the same no-parser feature set.

Target	Kind	Features / scope	Size
KiteSQL main c0e63a0 memory	embedded database	orm,macros; parser not feature-gated on main	7.25 MiB
KiteSQL PR 4ee5c66 memory	embedded database	orm; parser disabled	2.25 MiB
KiteSQL main c0e63a0 RocksDB	embedded database	orm,macros,rocksdb; parser not feature-gated on main	15.74 MiB
KiteSQL PR 4ee5c66 RocksDB	embedded database	orm,rocksdb; parser disabled	10.08 MiB
KiteSQL main c0e63a0 LMDB	embedded database	orm,macros,lmdb; parser not feature-gated on main	8.27 MiB
KiteSQL PR 4ee5c66 LMDB	embedded database	orm,lmdb; parser disabled	2.60 MiB
rusqlite 0.34.0 + bundled SQLite	SQLite binding / embedded SQLite	minimal in-memory SQL smoke app	1.31 MiB
SeaORM 2.0.0-rc.38	ORM framework, not a database	minimal SQLite query-builder/model app, no DB engine linked	0.39 MiB
Diesel 2.3.10 + bundled SQLite	ORM/query builder, not a database	minimal SQLite ORM app with bundled libsqlite3	1.36 MiB
Turso	embedded database	0.7.0-pre.7 and pinned 0.5.0 did not compile on this rustc due transitive/proc-macro API requirements	n/a

KiteSQL PR reduction vs main in this measurement:

• memory: 7.25 MiB -> 2.25 MiB, down 68.9%
• RocksDB: 15.74 MiB -> 10.08 MiB, down 36.0%
• LMDB: 8.27 MiB -> 2.60 MiB, down 68.5%

TPCC Throughput Comparison

Retested on the current local PR working tree after rebuilding target/release/tpcc without profiling features and clearing old TPCC database state for each backend. Each row is a fresh 1-warehouse run using TPCC's default 720-second measurement window. Raw logs were generated locally under tpcc/results/2026-06-14_720-full-current/ and are not committed. All runs passed TPCC constraint checks.

Backend	Profile	TpmC	New-Order p90	Payment p90	Order-Status p90	Delivery p90	Stock-Level p90	Measure Time	Status
KiteSQL LMDB	embedded database / LMDB	61723	0.001s	0.001s	0.001s	0.002s	0.001s	720s	ok
KiteSQL RocksDB	embedded database / RocksDB	30446	0.001s	0.001s	0.001s	0.016s	0.002s	720s	ok
SQLite balanced	embedded SQLite / balanced profile	42989	0.001s	0.001s	0.001s	0.001s	0.001s	720s	ok
SQLite practical	embedded SQLite / practical profile	42276	0.001s	0.001s	0.001s	0.001s	0.001s	720s	ok

These values are now mirrored in README.md and tpcc/README.md. The LMDB result is lower than the historical README value on the same machine, while RocksDB remains close to the historical value; the current run was kept as the source of truth because it is a full 720s matrix from clean databases.

Testing

• cargo test --workspace
• make wasm-build && make wasm-examples
• Size smoke builds for memory/RocksDB/LMDB configurations with parser disabled
• Full 720s TPCC matrix for KiteSQL LMDB, KiteSQL RocksDB, SQLite balanced, and SQLite practical