[JingsongLi]

Purpose

Closes #1303.

Common aggregate spellings, JSON output expressions, custom LIKE escapes and several SQL dialect entry points currently fail during SQLContext planning or execution. This change executes those patterns through existing DataFusion/Paimon primitives while preserving NULL, ordering, name-resolution and schema-evolution contracts.

Brief change log

• Add count_if, any, any_value, first, approx_count_distinct, max_by, min_by, arg_max, collect_list and LISTAGG ordered-set compatibility. Preserve filters, output names and registered-session-function precedence, and keep original REST definitions portable.
• Add typed to_json/json_object output and exact-Variant string CAST/TRY_CAST rewriting, including dictionary value NULLs, nested values, decimals and corrupted Variant handling.
• Add one-character LIKE/ILIKE escapes, assert_true, 64-bit conv, date_diff overloads and explicit-zone convert_tz.
• Route scalar MODIFY COLUMN through existing SchemaChange validation, protecting field IDs, unmentioned attributes and temporary-table shadowing. Add extended Paimon descriptions with native/temporary/external-object fallbacks and redacted credential-like properties.
• Keep extended descriptions in extended_describe and expression lowering in expression_compat. Selection aggregates adapt native first_value state and grouped accumulation while evaluating each value/key once.

Tests

• Added thirteen integration regression groups and one REST function/view unit regression. Reproduced the five main failure categories before implementation.
• Cover empty/all-NULL groups, OR filters, constant sort keys, duplicates/order, window counts, local UDF overrides, persistent definition round trips, dictionary value NULLs, malformed Variant CAST/TRY_CAST, decimal/binary JSON, Unicode LIKE, numeric/timezone/DST boundaries, actual type widening and schema-update failure atomicity.
• Passed final-head full validation: TZ=UTC RUST_MIN_STACK=16777216 CARGO_BUILD_JOBS=3 cargo test --locked -p paimon-datafusion --all-targets -- --test-threads=4 with Rust 1.95.0.
• Passed Rust 1.98 Clippy for the crate and all targets with --no-deps -- -D warnings, formatting and diff checks.
• Completed additional independent review after the module rename. Added failing regressions for invalid fixed-offset minutes, repeated evaluation of volatile aggregate arguments, nullable partial key state, dictionary value NULLs and nullable FILTER masks, then repaired them. Also cover native selection-state merging across batches/partitions and negative CONV source bases returning NULL.

API and Format

Adds SQL names and syntax. New bare compatibility names follow built-in precedence; explicitly registered session functions and qualified catalog calls retain their precedence. Extended Paimon DESCRIBE returns comment/metadata rows, while native and temporary objects retain their existing description behavior. No storage-format, dependency or Rust public API changes.

The documented initial scope uses native fixed-precision HLL, two-argument selection aggregates and scalar MODIFY types. Unsupported precision requests, selection window/DISTINCT modifiers, LISTAGG overflow clauses and nested MODIFY targets fail explicitly. Existing nullability-change policy remains in force.

Design scope

Paimon metadata, Variant conversion and schema changes stay in the integration layer. General expressions reuse native aggregators, Arrow values/timezones and existing Variant decoding, retaining native aggregate state layouts and existing storage contracts.

These are documented SQLContext compatibility subsets. Two-argument date_diff uses end/start date subtraction; the unit/start/end form has explicit UTC/calendar rules. JSON output retains nested NULLs and Arrow temporal formatting and does not implement Spark JSON options. CONV accepts source bases 2 through 36, with negative destination bases for signed output. Native HLL precision and explicit DST-ambiguity behavior are retained.

DataFusion 54's separate Spark package already covers some overlapping names, including collect_list and two-argument date_diff. This PR keeps small native adapters without adding that dependency or registering its full set of overriding functions; future expansion should evaluate selective upstream reuse first. Full Spark/MySQL/Trino emulation is outside this change.

Documentation

Updated docs/src/sql.md with syntax, supported inputs, NULL/error behavior, timezone/encoding rules, precedence and compatibility limits.

[leaves12138]

LGTM