Plan analysis: Bare Scan rule and Serial Plan benefit formula (#4526, #4528)

[erikdarlingdata]

Fixes #4526
Fixes #4528
Part of #4511

Why

PerformanceStudio has two small plan-analysis improvements that PerformanceMonitor's PerformanceMonitor.PlanAnalysis project hadn't picked up yet: a new rule for a full-table scan with no predicate, and a corrected formula for the Serial Plan benefit estimate.

What changes

Rule 34, "Bare Scan" (PlanAnalyzer.cs): flags a Clustered Index Scan or heap Table Scan that has no predicate and is not a lookup.

• An actual plan fires when the scan's own elapsed time is above 0ms; an estimated plan fires when the scan costs at least 20% of the plan.
• Three or fewer output columns suggests a covering nonclustered index. More than three columns suggests a columnstore index instead, because column count doesn't make a columnstore index more expensive the way it does a rowstore index.
• Added to BenefitScorer.OperatorTimeRules so it gets a benefit score, the same as the other scan rules.

Serial Plan benefit formula (BenefitScorer.cs): replaces the two-branch estimate (a flat 75% when CPU-bound, otherwise the CPU/elapsed ratio times 75% capped at 50%) with a single formula: (cpu * (DOP - 1) / DOP) / elapsed * 100, assuming a DOP of 4, capped at 100%. A statement whose subtree cost is under 1 gets no benefit at all, since a trivial plan doesn't gain from parallelism.

Both changes mirror PerformanceStudio's own commits behavior-for-behavior, adapted only for the difference that this project has no AnalyzerConfig, so there is no rule-disable guard.

Test plan

New class: Darling/Darling.Tests/PlanSync4526And4528Tests.cs, two test classes:

• PlanSync4526Tests: a bare heap scan with an actual own time and 2 output columns gets the covering-index advice; a bare clustered-index scan with 5 output columns gets the columnstore advice; a scan with a predicate gets no warning; an estimated plan at 25% of cost warns, one at 10% does not.

• PlanSync4528Tests: cpu 900 / elapsed 1000 / cost 5 scores 67.5% (the old formula capped this case at 50%); cpu 2000 / elapsed 1000 / cost 5 scores 100% (the old formula gave a flat 75%); a statement with subtree cost 0.5 gets no benefit at all (the old formula had no such gate).

• Runtime RED on dev: 6 of the 8 new facts fail there (the two "no warning" facts pass on dev, as they should): both Bare Scan warn facts and the estimated-plan fact, and all three Serial Plan benefit facts.

• Mutations, each reverted: the narrow-output threshold colCount <= 3 → <= 0 fails the covering-index fact; removing the StatementSubTreeCost >= 1.0 gate fails the trivial-cost fact.

• Run: the new classes plus the plan-analysis classes: Total: 239, Failed: 0, Skipped: 2 (the live-Postgres plan-tool class, which runs in CI). Lite.Tests builds. No test enumerates warning types, so no census pin changed.

Note: this PR also ports PerformanceStudio's single Serial Plan benefit formula into BenefitScorer. No entry point calls the scorer yet (#4546), so that part has no visible effect until it does. The user-facing entry for benefit values belongs with that change.

CHANGELOG

SECTION: Added
ENTRY: - Plan analysis: added the "Bare Scan" finding for a full-table or full-clustered-index scan with no predicate ([#4545]) - A Clustered Index Scan or heap Table Scan with no predicate now gets a covering-index or columnstore suggestion depending on its output width.
REF: [#4545]: #4545