Stage index_object_stats definition metadata instead of correlating per index (#3508)

[erikdarlingdata]

Closes #3508. The body-cost half of #3477 (the width half is the per-collector DB scope; this is the per-database body).

Problem

index_object_stats staged the three stats DMVs into #temps (the #1135 fix) but built the per-index definition metadata in the final SELECT as five correlated subqueries against the live catalog, per index row: key_columns/included_columns (FOR XML over sys.index_columns/sys.columns), two FK EXISTS over sys.foreign_key_columns, and the compression TOP (1) over sys.partitions. ~5 catalog probes per index, scaling with index count — a 54K-index database spends ~28 s in this body.

Fix

Extend the existing staging to the definition catalog — the technique sp_IndexCleanup already uses. Stage sys.index_columns (+ column names, once, with a supporting clustered index), the FK-backed / FK-referenced key-column sets, and the per-index lowest partition compression into #temps; then LEFT JOIN them in the final SELECT with no correlated subqueries. Emitted strings and flags are byte-identical, so Stage-2 dedupe and FK-protection are unaffected — a plan-shape change only. The collector is shared, so Lite and Darling both get it.

Measured

Synthetic database, 56,001 indexes on empty tables (so this isolates the definition-metadata cost):

	body time
before (correlated)	~4.1 s
after (staged)	~2.0 s

EXCEPT both ways returns 0 rows in either direction at both 12K and 56K indexes. Per-step instrumentation confirms the correlated per-index catalog probing is gone; the residual is the unchanged dm_db_index_operational_stats read plus the irreducible cost of building the delimited key/include strings. (Empty tables, so absolute times run below a real 54K-index instance whose operational-stats DMV is heavier; the ~2x and the parity are the transferable results.)

Tests

• New guard DefinitionMetadata_IsStaged_NotCorrelatedPerIndex pins that the definition metadata is staged, not correlated (a revert drops the markers).
• Full local suites green: Lite.Tests 3916/0, Darling.Tests 9627 total / 0 errors. The one Darling failure (NpgsqlRootCertificateValidationTests.ChainShape…) is the known machine-local TLS cert-chain test on a path this change never touches (it passes in CI on Linux).

🤖 Generated with Claude Code

https://claude.ai/code/session_01RzwoqcD62jKyp6tzsWpZLz