SQL lowering: run DataFusion's SimplifyExpressions (needs IR timestamp literals and regex operators) #650

Problem

The SQL front end lowers DataFusion's plan right after type coercion and does not run DataFusion's expression simplification (SimplifyExpressions). A DataFusion query reaches its executor simplified, and ours does not, so the lowered IR can differ from what DataFusion would execute. Today only one piece is matched by hand: #539 folds a numeric cast of a literal (latency < CAST(100 AS DOUBLE) → latency < 100.0) with DataFusion's ScalarValue::cast_to.

Running the whole rule (tried on the #646 restack, 2026-10-10) breaks the SQL corpora (bgp_analytics, bgp_jan2024_workload, netflow, tpch_deequ) in two ways:

1. Timestamp literals. time >= CAST('2018-05-17 13:10:01' AS TIMESTAMP) folds to TimestampMillisecond(1526562601000). The IR has no timestamp literal, so lowering fails with unsupported scalar. Nearly every BGP and NetFlow query has a time filter.
2. Regex. REGEXP_LIKE(col, pat) is rewritten into the operator col ~ pat (RegexMatch), which lowering does not support (unsupported feature: operator: RegexMatch).

The rule rewrites more than these two; the corpora only hit these.

Proposal

Run DataFusion's SimplifyExpressions after TypeCoercion, with an OptimizerContext without a query start time (without_query_execution_start_time()), so now() stays symbolic and still lowers to CurrentTimestamp. To make that work:

• Add a timestamp literal to the IR's ScalarValue (types, planner, executor).
• Lower RegexMatch / RegexIMatch / RegexNotMatch / RegexNotIMatch, and any other shape the simplifier produces, found by running the corpora.
• Drop the hand-written numeric cast fold from feat: lower SQL through unified operators and scalars #539.

A timestamp literal also lets summary coverage put an absolute time filter (ts >= t0 AND ts < t1) into the selection, which #573 §4.2.2 lists as "later": today it stays in the definition.

Related

#539 (SQL lowering), #646 / #573 (summary coverage).