[milindsrivastava1997]

Summary

• SQLWindowFunc previously dropped the window-frame clause entirely. A default frame, ROWS BETWEEN 2 PRECEDING AND CURRENT ROW, and ROWS BETWEEN CURRENT ROW AND UNBOUNDED FOLLOWING therefore lowered to structurally identical IR.
• This PR adds first-class WindowFrame, WindowFrameUnits, WindowFrameBound, and typed WindowFrameOffset values to the canonical IR.
• lower_window now preserves DataFusion's resolved frame, including its context-dependent default.
• Offset RANGE ... INTERVAL bounds are normalized to month/day/nanosecond components instead of remaining opaque DataFusion 43 strings such as "1 HOUR".
• Legacy serialized SQLWindowFunc nodes that predate the frame field remain readable through #[serde(default)] frame: Option<WindowFrame>; newly lowered SQL always produces Some(frame).

These defaults follow DataFusion's documented behavior: with ORDER BY, the default ends at the current row; without ORDER BY, it covers the entire partition. None is reserved for legacy IR because one static fallback could not reconstruct both cases safely.

Before / after IR

Example:

SELECT service, SUM(latency) OVER (
  PARTITION BY service ORDER BY ts
  ROWS BETWEEN 2 PRECEDING AND CURRENT ROW
) FROM metrics

Before — no machine-readable frame:

SQLWindowFunc {
    func: Sum,
    args: [Column(2)],
    partition_by: GroupKeys { keys: [1], without: false },
    order_by: [SortKey { expr: Column(0), ascending: true, nulls_first: false }],
    output_name: "... ROWS BETWEEN 2 PRECEDING AND CURRENT ROW",
    child: Scan { .. },
}

After — the resolved frame is part of the IR:

SQLWindowFunc {
    func: Sum,
    args: [Column(2)],
    partition_by: GroupKeys { keys: [1], without: false },
    order_by: [SortKey { expr: Column(0), ascending: true, nulls_first: false }],
    frame: Some(WindowFrame {
        units: Rows,
        start_bound: Preceding(Scalar(Int64(2))),
        end_bound: CurrentRow,
    }),
    output_name: "... ROWS BETWEEN 2 PRECEDING AND CURRENT ROW",
    child: Scan { .. },
}

An interval RANGE bound is typed as:

Preceding(Interval {
    months: 0,
    days: 0,
    nanoseconds: 3_600_000_000_000,
})

for INTERVAL '1' HOUR.

Scope

• ROWS and RANGE are modeled.
• GROUPS is explicitly rejected rather than silently mis-lowered. DataFusion supports it, but it remains outside this PR's declared scope.
• This PR preserves frame semantics for downstream consumers; it does not add a post-ASAP frame-execution rule.

Tests

• window_frame_is_captured_not_dropped
• range_interval_frame_is_preserved
• groups_frame_is_rejected
• legacy_window_json_without_frame_deserializes_as_unspecified
• cargo test -p asap-frontend-sql -p asap-types
• CI format-and-lint and test

Fixes #268

[zzylol]

Fixed the review findings in commit 724279d.

What changed:

• Added a typed WindowFrameOffset representation. DataFusion 43 exposes SQL frame interval literals such as INTERVAL '1' HOUR as normalized UTF-8 text ("1 HOUR"), so lowering now parses that text with Arrow’s interval parser and stores canonical month/day/nanosecond components. Native DataFusion interval scalar variants are normalized into the same representation.
• Changed SQLWindowFunc.frame to Option<WindowFrame> with #[serde(default)]. Newly lowered SQL always writes Some(frame); legacy serialized IR without the field now reads as None, preserving the fact that the old planner did not retain enough context to reconstruct its implicit frame safely.
• Added regression tests for an offset RANGE BETWEEN INTERVAL '1' HOUR PRECEDING ... frame and for deserializing legacy frame-less window JSON.

Validation: cargo test -p asap-frontend-sql -p asap-types passes (69 SQL lowering tests and 99 asap-types tests, plus the corpus/unit suites).