[erikdarlingdata]

Fixes #4611. Also carries #4605's rollup materialization ceiling (RollupCoverage.CeilingOf), which this uses.

Summary

get_query_store_duration_trend and the desktop Query Store trend chart used to run min(bucket)/max(bucket) over query_store_stats_corrected_hourly on every call, to decide where the rollup-served region ends and the raw tail begins. Both callers already probe and cache a RollupCoverage for the whole store on a timer, and that coverage now carries each rollup's measured materialization ceiling. This change reads the floor and ceiling for the corrected hourly rollup out of that cache instead of issuing a fresh bounds query on every trend load.

The ceiling (#4605)

RollupCoverage gains a per-view materialization CEILING: the instant at or above which a rollup holds nothing yet, read from the engine's own watermark (or a max(bucket) + bucket width fallback) in the same probe cycle that measures the floor, and cached the same way. CeilingOf answers null for a view that was never probed, is absent, or has materialized nothing — the safe "unknown" reading that keeps a caller on the raw path. Its own pins: an unknown-view pin, a cache-merge pin, and a live test against a real TimescaleDB (RollupCoverageCeilingLiveTests) proving the ceiling matches the engine's watermark, advances on a further refresh, and holds while the cache's reuse window is still open.

What changed

• QueryStoreTrendRouting gets a ResolveAsync(RollupCoverage, NpgsqlDataSource, CancellationToken) overload. When the coverage carries a ceiling for query_store_stats_corrected_hourly, it builds the route directly from CeilingOf/FloorOf — no store round trip. CeilingOf is already the exact exclusive boundary the old pure decision computed as newest bucket + bucket width, so it becomes RawStartUtc with no adjustment.
• When the coverage has no ceiling for the view (never probed, absent, or nothing materialized), it falls back to the existing ResolveAsync(NpgsqlDataSource, ...) probe-and-read path unchanged, so a cold or unknown coverage is never less correct than before.
• The MCP tool (DarlingMcpTrendTools) and the desktop viewer (ViewerDataService.QueryTrends) both now fetch their already-cached coverage first and pass it in, instead of letting the resolver probe on its own.
• Routing decisions and the raw-only fallback SQL are unchanged.

Test plan

Run on a local TimescaleDB rig (timescale/timescaledb:2.30.1-pg18, shared_preload_libraries=timescaledb,pg_stat_statements):

• New live pin, QueryStoreTrendRoutingCachedLiveTests.ResolveAsync_WithACachedCeiling_IssuesNoBoundsProbe_AndMatchesTheFallbackRoute: seeds two hours of Query Store snapshots, materializes the rollup, resolves the route once through the uncached probe-and-read overload as the reference answer, then resolves it again through the coverage-taking overload and asserts the two routes agree (UseRollup, RawStartUtc, RollupFloorUtc) and that pg_stat_statements shows ZERO calls matching the min(bucket)/max(bucket) bounds statement after the cached call — the same pg_stat_statements counting discipline StoreSizeCacheLiveTests uses against the product's own call path, not a mock.
  • GREEN against this branch: Total: 1, Failed: 0.
  • RED: with the coverage-taking overload's ceiling short-circuit temporarily removed (falling straight to the uncached probe-and-read for every call, the pre-The Query Store trend router re-runs an uncached min/max(bucket) on the corrected hourly rollup for every get_query_trend call and trend chart load (~84 MB temp each) #4611 shape), the same pin fails: Total: 1, Failed: 1. Reverted after confirming; the branch is clean.
• QueryStoreTrendRoutingTests, QueryStoreTrendRoutingLiveTests, QueryStoreTrendRoutingCachedLiveTests, RollupCoverageCeilingLiveTests, RollupCoverageRoutingTests, StorageCommandTimeoutTests, StartupCommandTimeoutTests: Total: 98, Failed: 0.
• The test-census family (CommentFilterAdoptionTests, DocCommentHygieneTests, McpPayloadContractCensusTests, every *Census* class): Total: 355, Failed: 0.
• McpToolsListBudgetTests: Total: 5, Failed: 0.
• Build: Darling.Tests, the Darling Viewer, and the Darling Service each build with -p:EnableWindowsTargeting=true at 0 warnings, 0 errors.
• Lite.Tests: no Lite/DuckDB code changes here, so no Lite class is affected.

CHANGELOG

SECTION: Changed
ENTRY: - The Darling Viewer's Query Store trend chart and get_query_store_duration_trend no longer re-scan the hourly rollup's oldest and newest hour on every load ([#4618]) - They take the rollup's range from the coverage the service already caches, instead of a fresh min/max read over the rollup on every call.
REF: [#4618]: #4618