[SPARK-56949][SQL] DecimalAggregates drops evalMode when rewriting Sum/Average

[yaooqinn]

What changes were proposed in this pull request?

In Optimizer.DecimalAggregates, rewrite the 6 Sum / Average rewrites
from the single-arg helper ctor (Sum(child) / Average(child)) to
original.copy(child = UnscaledValue(...)), so sibling fields like
evalMode are preserved instead of re-read from SQLConf. The 6 sites
are 4 Aggregate arms (SUM/AVG x un-widened/widened peel) plus 2 Window
arms (un-widened SUM and AVG); both Window arms are directly reachable
because the input is a bare AttributeReference, not a hoisted
composite Cast.

The helper ctors themselves are retained — used by other call sites.

Why are the changes needed?

Latent bug since SPARK-38432 added evalMode to Sum / Average: the
helper ctor re-reads EvalMode from SQLConf, silently dropping
EvalMode.TRY from try_sum / try_avg. Sum.shouldTrackIsEmpty = evalMode == TRY then drops the isEmpty aggregation buffer column, so
overflow no longer returns NULL.

Does this PR introduce any user-facing change?

Yes. try_sum / try_avg on narrow-precision decimals now preserve
EvalMode.TRY through the optimizer rewrite, so overflow returns NULL
as documented instead of wrapping (ansi=false) or throwing
(ansi=true). Affected surface: SUM precision <= 8, AVG precision <= 7
(both Aggregate and Window arms).

This is observable directly in the optimized plan — the rewritten Sum /
Average node carries evalMode = TRY post-fix vs evalMode = ANSI (the
SQLConf default, re-read by the helper ctor) pre-fix.

How was this patch tested?

6 new unit tests in DecimalAggregatesSuite asserting evalMode == TRY
post-rewrite for: un-widened Aggregate SUM, un-widened Aggregate AVG,
widened-cast Aggregate SUM, widened-cast Aggregate AVG, un-widened
Window SUM, un-widened Window AVG.

• Pre-fix: 6/6 FAIL with Expected evalMode=TRY post-rewrite, got ANSI
• Post-fix: 37/37 PASS (6 new + 31 existing, 0 regression)
• Sibling regressions clean: DecimalPrecisionSuite +
  AggregateOptimizeSuite 20/20 PASS
• dev/lint-scala clean

Backport candidates: branch-3.5 and branch-4.0 carry the 4 helper-ctor
sites (2 Aggregate + 2 Window) — verified by git grep "aggregateFunction = (Sum|Average)\(UnscaledValue". The 2 widened-cast Aggregate sites are
master-only (SPARK-56627). Backport patches would touch only the 4 shared
sites.

Was this patch authored or co-authored using generative AI tooling?

Generated-by: Claude Opus 4.7

[peter-toth]

Does this bug affect the above case Sum(e @ DecimalExpression(prec, scale), _) and case Average(e @ DecimalExpression(prec, scale), _) branches?

[peter-toth]

@yaooqinn, I think the issue is real, but is the example in description valid?

scala> sql("SELECT try_sum(cast(id as decimal(7,2))) FROM range(10)").show
+---------------------------------+
|try_sum(CAST(id AS DECIMAL(7,2)))|
+---------------------------------+
|                            45.00|
+---------------------------------+

[yaooqinn]

Thanks @peter-toth. Two updates:

1. Description: dropped the misleading 10-row overflow example; rewrote
   to frame the change as evalMode preserved through the rewrite
   (observable on the optimized plan).

2. Scope: expanded 4 → 6 sites. The un-widened Window arms at
   L2589 / L2594 carry the same helper-ctor bug and are directly
   reachable (bare AttributeReference, no ExtractWindowExpressions hoist).
   Added 2 tests; pre-fix both fail with got ANSI, post-fix 37/37 pass.

Force-pushed to b95abfb.

[yaooqinn]

Merged to master and 4.x, thank you @peter-toth