feat(executor): execute the exact SQL entropy fallback; Example 2 status

[zzylol]

Rebased on main d4869a7 (DF 54).

Wave 1 chain: #594 → #593 → #592 → #591 → #595 → #596 → #597 → #598 → #599 (on #589)

Why

#598 recognizes the SQL entropy idiom only under narrow conditions. Any other form, such as a nullable src_ip, stays relational, and its native binding failed:

Piece	Before
SUM(COUNT(*)) OVER ()	value operation has no native implementation
LN(p)	expression validate rejects ln

The parked #565 added both on the old stack. This PR ports it and records where #509 Example 2 stands on the current Stage 1 to Stage 3 pipeline. Part of #509 / #580 (W8).

What

• Operator::sql_window_sum(input, column, name) appends one nullable total to every row (SUM semantics, NULL for empty or all-NULL input). It is memory-accounted and cooperative, and survives the UncheckedOperator round trip. Binding accepts only SUM(col) OVER () with the unbounded frame. Partitioned, ordered and finite frames return native SQL window SUM requires the complete unordered relation.
• SQL ln uses the same numeric path as sqrt.
• integration-tests/tests/planner_layering_example2.rs records Example 2's status.

After: the original Q2 SQL and its recognized form return the same values on every fixture: NULL, -0.0, ln 2, and -0.75 ln 0.75 - 0.25 ln 0.25.

Example 2 status (asserted by planner_layering_example2)

	Q1 distinct	Q2 entropy	Q3 L2
Lowers to the intent	Cardinality	FrequencyEntropy (the doc's text)	FrequencyL2 only with a Float64 product. The doc's SUM(c * c) is Int64 and stays relational
Time filter	WHERE ts >= now() - INTERVAL '1 minute' becomes a Scan predicate (CAST(ts) >= CAST(now() - 60 s)). It is not a per-measure FILTER and not a TimeRange, and Pass 1 accepts it	same	same
plan_stages	plans; Stage 1 offers PassThrough + UnivMon (+ HLL)	plans; PassThrough + UnivMon	plans; PassThrough + UnivMon
Exact execution	2	entropy in nats	√5

Remaining blockers for the shared UnivMon plan:

1. No accuracy model for UnivMon, so Stage 3 rejects every UnivMon candidate (no accuracy model for UnivMon). Another agent is adding this.
2. No summary-capability rule in Pass 2, so the three UnivMon builds are never merged. Another agent is adding this.
3. The time filter carries no lookback that Stage 3 or coverage can see: it is a scan predicate, not a TimeRange. The runtime also has no now(): binding the windowed queries fails with unsupported scalar expression: CurrentTimestamp, so the exact runs above drop the filter.
4. The design's integer Q3 is not recognized (overflow).
5. Unrelated to sharing: Stage 1 also offers an exact Count accumulator for each population guard, and it fails Stage 2 over SQL rows (ColumnRef::SampleValue has no value column).

Tests

Ported from #565: complete_sql_sum_window_preserves_rows_and_nulls, sql_ln_executes_numeric_and_null_arguments, complete_window_sum_enforces_workspace_budget, the original-vs-rewrite entropy comparison, and native_sql_sum_window_rejects_other_frames. Before the port, both entropy integration tests fail. The developer-doc additions from #565 are not ported.

Gate: fmt and clippy (-D warnings) pass. cargo test --workspace: 1,546 passed, 10 ignored (#589: 1,517 / 10).

Ports #565. Stacked on #598.

🤖 Generated with Claude Code