MCP top-queries/top-procedures disclose a shortened raw-tier window (#4231)

[erikdarlingdata]

Part of #4231.

Why

At "Last 7 days", the WPF Queries grids and two MCP tools read raw tables. Once the rollups are armed, TimescaleDB drops those tables' chunks at 4 days, and none of these reads said so. #2364 fixed this for get_query_store_top: it reads the window floor and reports effective_start and window_truncated.

get_top_queries_by_cpu and get_top_procedures_by_cpu, over query_stats and procedure_stats, never got that fix. A caller who asked for 7 days of CPU ranking got about 4 days back, and hours_back came back unchanged. The WPF grids showed the same short window with no note.

What changes

This is the Darling half of #4231 (stage 1 of its ruling). The Lite half is #4279. Stage 3, routing the top-N reads to the hourly rollups, waits on #4186.

• RawWindowFloor (new, in PerformanceMonitor.Darling.Storage) is one floor probe for the three raw tables: query_stats, procedure_stats and query_store_stats. It generalizes get_query_store_top reports a window it cannot serve: raw query_store_stats is dropped at 4 days #2364's DarlingDataReader.QueryStoreWindowFloorSql. It lives in Storage so the WPF viewer calls the same probe without referencing the service assembly. IsTruncated reuses DurationTrendRouting.TruncationSlack, the 90-minute boundary that get_query_store_top and the duration trends already use.
• DarlingDataReader: QueryStoreWindowFloorSql and GetQueryStoreWindowFloorAsync now call the shared probe, with the same SQL text and behavior. New GetQueryStatsWindowFloorAsync and GetProcedureStatsWindowFloorAsync serve the two tools.
• get_top_queries_by_cpu and get_top_procedures_by_cpu now return effective_start, effective_hours_back and window_truncated, plus a truncation_note when the window was cut short. The note text is Lite's, word for word (Lite MCP tools disclose a truncated query-stats window (#4231) #4279). The names and meaning match get_query_store_top. Every existing field, default and argument is unchanged.
• get_query_store_top calls RawWindowFloor.IsTruncated in place of its own AddMinutes(90). The value is the same.
• WPF viewer: the Top Queries, Top Procedures and Query Store grids each read their table's floor through the shared probe. When the floor cuts the window short, a note above the grid says "Showing since" and the effective start. The time is in the viewer's display time, formatted like the trend chart titles. The note refreshes on a toolbar-window load and on each slicer drag. Every window it probes is UTC, the same window the grid read.
• Web viewer: the three tools' truncation_note shows through the existing get_pg_index_bloat needs a coverage summary (trusted/suppressed counts) — its unmeasured-first sort has misled three sessions in one night #3278 noteKey option, on the built-in server tabs and in the starter dashboard template. The Top Queries panel on the Queries tab is built by hand, so it renders the note itself.
• Tests and pins: rows for the two tools in McpPayloadContractCensusTests (WindowFloorBlocks, WindowFloorTools), and the tools/list byte pins (DarlingMcpDataTools.txt, McpToolsListBudgetTests).

The probe's cost

On a rig with 6 daily chunks per table for one server, EXPLAIN (ANALYZE, BUFFERS) showed an index-only scan under a Limit for each table:

Table	Index	Buffers	Time
query_stats	idx_query_stats_time	5	0.054 ms
procedure_stats	idx_procedure_stats_time	2	0.024 ms
query_store_stats	idx_query_store_stats_time	2	0.034 ms

Each read stops at the first row of the oldest chunk in the window, even with 6 chunks present.

The served descriptions

The two tools' served heads are unchanged from dev (474 and 406 bytes). Both tools are shared with Lite. McpToolGuideTests.EverySharedToolName_CarriesTheMarkerOnBothSkus_OrNeither_WithByteIdenticalHeads requires their heads to match Lite's byte for byte, so this PR cannot change them alone. The disclosure sits in each tool's get_tool_guide tail, through McpHelpers.WindowTruncatedDescription. It comes before McpToolGuideTopics.CpuTimeExtremesAndAttribution, so the tail still ends on the CPU topic that McpToolGuideHeadsDataTests.CpuTimeExtremesTopic_RidesOnBothTopByCpuTools requires.

For the same reason, McpPayloadContractCensusTests.EveryWindowFloorTool_CarriesTheSharedClause_AndNoOtherToolDoes reads these two tools' guardrail from the full description, not the head. The code states that reason.

TotalCeilingBytes is 174,373, measured on the merged tree after #4272 and #4273. This PR adds no head bytes.

Left for #4231

• Lite MCP tools disclose a truncated query-stats window (#4231) #4279, the Lite half, merges after this PR. Its last step updates the shared heads in both products at once. get_query_store_top's "Lite: no such floor" line becomes false once Lite has the floor. The same step moves the window guardrail into the two CPU tools' heads and drops the census carve-out above.
• Stage 3: routing the top-N reads to the hourly rollups, after Legacy trio off the refresh grid (#3653) #4186 merges.

Test plan

• dotnet build Darling/Darling.Tests/Darling.Tests.csproj -c Debug: 0 warnings, 0 errors.
• Full Darling.Tests, once, on commit e67ff186 after a merge of dev: 13,974 total, 5 failed, 49 skipped. The lane fixed all five, and two more that the fixes caused:
  • RepoFileAdoptionTests.TheLfReadingPins_AreExactlyTheOnesDeclaredHere: RawWindowFloorViewerPortTests.cs joins the declared LF-read list, because its web-source pin counts noteKey: "truncation_note" exactly only under an LF read.
  • McpToolGuideHeadsDataTests.EveryConvertedHead_CarriesItsGuardrailFact_AndThePointer and McpToolGuideTests.EveryConvertedHead_StaysAtOrUnder620Characters: the two heads are back at dev's length.
  • McpToolGuideHeadsDataTests.CpuTimeExtremesTopic_RidesOnBothTopByCpuTools: the tail order is swapped, as above.
  • McpToolGuideTests.EverySharedToolName_CarriesTheMarkerOnBothSkus_OrNeither_WithByteIdenticalHeads: both heads match Lite's again.
  • McpPayloadContractCensusTests.EveryWindowFloorTool_CarriesTheSharedClause_AndNoOtherToolDoes: the carve-out above.
  • McpToolsListBudgetTests.EveryServedDescription_AndEveryParameterDescription_MatchesItsCeiling: DarlingMcpDataTools.txt holds the measured 406 and 474 bytes.
• The PR tender merged dev with MCP budget: describe_custom_view_catalog groups measures by source, under the response budget (#4198) #4272 and MCP budget: get_query_store_top's default stays under 32 KB (#4198) #4273 (commit 40f58504). Then these classes passed with no failures: McpToolsListBudgetTests, McpPayloadContractCensusTests, McpToolGuideTests, McpToolGuideHeadsDataTests, RepoFileAdoptionTests, the QueryStoreTop and QueryStoreClutter classes, the RawWindowFloor classes, QueryStoreRegressionsWebDefaultTests and DarlingWebEndpointsTests.
• The live classes (RawWindowFloorMcpLiveTests and the live DarlingMcpDataToolsTests) need PostgreSQL. They skipped locally after the last fixes, so GitHub Actions CI decides.

CHANGELOG entry

SECTION: Fixed
ENTRY:

• Top-CPU reads disclose a short window ([MCP top-queries/top-procedures disclose a shortened raw-tier window (#4231) #4278]) - Two MCP tools read raw tables that the store purges at 4 days with the rollups on. They are get_top_queries_by_cpu and get_top_procedures_by_cpu. A longer window came back short, and hours_back came back unchanged. Both tools now return effective_start, effective_hours_back and window_truncated, plus a truncation_note, like get_query_store_top. The desktop viewer's Top Queries, Top Procedures and Query Store grids show "Showing since" and the effective start in the same case. The web viewer shows the same note on those panels.
  REF:
  [MCP top-queries/top-procedures disclose a shortened raw-tier window (#4231) #4278]: MCP top-queries/top-procedures disclose a shortened raw-tier window (#4231) #4278

🤖 Generated with Claude Code

https://claude.ai/code/session_01FVjn4PBJN71NQXdFo6ZxNQ