feat: add scalable SQL execution paths

[HelgeSverre]

Summary

• add scalable SQL execution paths for composite index ranges, correlated EXISTS, selective indexed joins, spill-backed DISTINCT, incremental window aggregation, bulk loading, and populated-table ADD PRIMARY KEY
• preserve atomic DDL rewrites with serializable validation while bounding rewrite memory
• fix exact window aggregation above 2^53 and avoid floating-point prefix cancellation
• make large DISTINCT byte-bounded, spill-capable, cancellation-aware, and safe for async runtime workers
• preserve mixed-numeric DISTINCT equality and create result spill files with owner-only permissions on Unix

Review follow-up

The adversarial review identified correctness and resource-safety gaps in the original implementation. This update adds regression coverage and fixes for:

• lossy f64 prefix subtraction in window SUM/AVG
• unbounded memory during ALTER TABLE ... ADD PRIMARY KEY
• row-count-only DISTINCT memory accounting
• synchronous DISTINCT spill work on Tokio workers
• mixed numeric representations surviving DISTINCT
• overly permissive DISTINCT result spill files

Implicit ALTER transactions now skip unnecessary per-key checkpoint logging because the entire private transaction can be discarded on failure. Explicit user transactions retain checkpoints and statement rollback semantics.

Performance

Release builds on the same Apple Silicon host, identical schemas/data, alternating execution order, five warmups, and 20-100 measured samples. Values are median client-observed latency.

Review fixes: before vs after

Workload	Before	After	Change
DISTINCT spill, 50k rows / 1k groups	28.58 ms	28.27 ms	-1.1%
Integer RANGE window, 5k rows	4.36 ms	4.39 ms	+0.7%
Float cumulative window, 10k rows	3.29 ms	3.67 ms	+11.6%
ADD PRIMARY KEY, 20k rows + secondary index	216.36 ms	208.75 ms	-3.5%

The floating-window regression test initially exposed a 3.66 ms to 2,378 ms slowdown from frame rescanning. A range-sum tree restored bounded evaluation; the final 3.67 ms result retains exact singleton-frame behavior.

ElyraSQL vs MySQL 8.0.33

Workload	MySQL	ElyraSQL	Relative
PK point lookup	0.111 ms	0.140 ms	ElyraSQL 26% slower
DISTINCT, 50k rows / 1k groups	10.50 ms	14.22 ms	ElyraSQL 35% slower
Integer RANGE window, 5k rows	4.20 ms	3.87 ms	ElyraSQL 8% faster
Float cumulative window, 10k rows	3.34 ms	3.65 ms	ElyraSQL 9% slower
ADD PRIMARY KEY, 20k rows + secondary index	45.34 ms	208.75 ms	ElyraSQL 4.6x slower

Profiling shows the remaining ADD PRIMARY KEY gap is dominated by atomic redb validation and individual key removal/insertion. A storage-level atomic bulk/range-rewrite primitive is intentionally deferred to a follow-up PR.

Verification

• cargo test --workspace --all-targets — 407 tests passed
• cargo clippy --workspace --all-targets -- -D warnings
• cargo fmt --all -- --check
• git diff --check
• targeted wire tests for ALTER atomicity, concurrent insert rejection, and DISTINCT behavior