Top queries and procedures over long windows drop filters, report zeros for missing columns, and overstate coverage #4711

When get_top_queries_by_cpu or get_top_procedures_by_cpu asks for a window longer than raw retention (about 72 hours and up), the read switches to the hourly rollup.

• That path ignores parallel_only, min_dop and group_by=host_object, but the payload reports them as applied, and an empty page says the filter was applied in SQL.
• Columns the rollup does not carry come back as 0 instead of null: reads, writes, physical reads, rows, spills, DOP and plan hash. The payload then shows is_parallel=false and avg_reads=0, although zero means "measured zero" everywhere else.
• min_cpu_ms and max_cpu_ms on this path are the min and max of per-collection deltas, not the per-execution extremes the tool text promises.
• Elapsed min/max and sql_handle exist in the rollup, but the query reader doesn't select them.
• Query text is looked up one row at a time with no time bound, and a miss returns empty text with no note.
• Coverage is reported as complete without checking where the rollup starts. A young store, or a young server on an older store, reports window_truncated=false for a window the data doesn't cover.
• The hourly predicate drops the partial hour at the window start and includes the whole hour at an as_of end.

Evidence (1248a67)

Paths under Darling/PerformanceMonitor.Darling.Service/Mcp/ unless noted:

• DarlingMcpDataTools.cs:566-569,575-578,584-586,601-612,645-661,694-701; procedures 754-756,769-773,795,810-814,822-823,837-841.
• DarlingDataReader.cs:1220-1254,1262-1271,1312-1329,1413-1442,1519-1524,1649.
• Darling/PerformanceMonitor.Darling.Storage/RetentionTierRouter.cs:68,210-260; rollup columns in TimescaleSupport.cs:2094-2098,9261,13510-13512.
• The repo's rule: DarlingMcpHostService.cs:990-992 (a silently dropped knob is a confidently wrong answer). The live test doc Darling.Tests/TopQueriesHourlyRoutingLiveTests.cs:180-185 says the note should say the filter was NOT applied, but the test only checks that the note is non-empty.

Fix expected

1. Filters: pass minMaxDop and rollUp into the router and force the raw tier when either is set (its floor probe already reports window_truncated). Refusing on the hourly tier doesn't work, because routing happens inside the reader. Make filter_applied, the group_by echo and the empty-page text tier-aware.
2. Columns: null for reads, writes, physical reads, rows, spills, DOP and plan hash, and null min/max_cpu_ms on the hourly tier. Select elapsed min/max and sql_handle (the procedures reader already does). object_type at DarlingMcpDataTools.cs:795 is the in-file precedent.
3. Text: one LEFT JOIN LATERAL in the ranked statement (like TopQueriesSql); null text plus a note when no raw row holds the hash. "Bounded by the raw floor" is a no-op.
4. Coverage: a per-server first-bucket probe over the same stitched relation, judged with RawWindowFloor.IsTruncated (its 90-minute slack covers a first bucket up to an hour late). The store-wide floor alone misses the young-server case.
5. Edges: snap effective_start to the first bucket and add a note on the trailing edge. The same predicate is used across the trend readers, so disclose rather than change one tool's predicate.

Tests that must fail before the fix

• An hourly-tier call with min_dop: today it returns a rollup payload that claims the filter applied.
• Missing columns are null, not 0, on the hourly tier.
• A server whose first bucket is after the window start reports window_truncated=true.

Constraints

• Run the MCP census and pin tests (McpPayloadContractCensusTests, the inventory pins).