The trend family's window floor is window_truncated on both SKUs, not the page dialect's truncated: get_query_trend, the duration-trend trio and get_query_store_top publish the store's reach under its own key, every consumer and description follows, and the vocabulary census holds the two facts apart (partial #3653 — item 17)

[erikdarlingdata]

One key carried two facts with opposite remedies, and the trend family's was the wrong one to be wearing the page dialect's spelling

Partial #3653 — item 17 (A15/A16), the clause the item's own text marks as NOT done: "window_truncated NOT done — ~15 consumers incl. get_query_store_top in DarlingMcpDataTools.cs, counted on the PR." #3703 found the homonym and classified it as stated residue; #3706 measured the blast radius, found it past the brief's ceiling, and stopped with a count. This is the rename, on both SKUs, with the census that keeps it renamed.

The mechanism

truncated on the MCP wire meant two things. On twenty-odd paged tools it is the page cut: the caller's limit bit, observed off a cap + 1 fetch (McpHelpers.BoundPage), published beside a *_returned count, and the remedy is a bigger limit. On the #2364 / #2353 trend family — get_query_trend, get_query_duration_trend, get_procedure_duration_trend, get_query_store_duration_trend — and on get_query_store_top, the same key was the window floor: the served series begins later than the requested start because the tier that answered no longer holds the window's head (raw dropped at four days with the rollups armed; a rollup that has not materialized the head; a store younger than the window; on Lite a retention_days purge or a server added mid-window). It rides beside effective_start / effective_hours_back, is judged off the served head against the shared ninety-minute TruncationSlack, and no limit changes it — get_query_store_top's top is a rank, and the trend tools have no cap at all. A client that had learned the first meaning read "raise the limit" off a fact no limit changes.

The window floor is now window_truncated — the <bound>_truncated dialect #3703 classified for a second bound in one payload (scan_truncated was its first member). The page cut keeps truncated. Nothing else in the disclosure block moves: the six keys are still six, in the same order, on the data path and the empty one.

What moved (the count #3706 asked for)

Before this PR, the window-floor fact under bare truncated had 4 emitters, 1 web consumer, 7 instructions rows naming the key, 0 web-catalog rows naming it, and 17 test asserts reading it. After: 0 window-floor sites spell it truncated; window_truncated has 4 emitters, 1 consumer, 9 instructions rows, 5 catalog rows, 9 tool descriptions and the same 17 asserts.

surface	site	what changed
emitter	DarlingMcpTrendTools.GetQueryTrend initializer	window_truncated = history.Truncated
emitter	DarlingMcpTrendTools.TrendDisclosure.WriteTo	envelope["window_truncated"] — the trio's data AND empty envelopes
emitter	DarlingMcpDataTools.GetQueryStoreTop initializer	window_truncated = truncated; truncation_note keeps its name (it is the prose for THIS flag; *_note is the idiom)
emitter	Lite McpQueryTools.WriteDisclosure	envelope["window_truncated"] — the trio + Lite's get_query_trend
consumer	server-tabs.js drawQueryTrend	reads trend.data.window_truncated; the notice sentence is word for word. Reading the old key would have failed SILENTLY — undefined is falsy, so the "history starts N hours back" notice would simply stop appearing under an axis that says otherwise
description	9 tools (Darling 5, Lite 4)	McpHelpers.WindowTruncatedDescription, one shared constant appended to each: names the key, the reach beside it, that it is not a page cut, and the WIRE CHANGE ("spelled truncated before #3653; the old spelling is not published here any more") — the #3719 shape, per tool this time because every one of these descriptions is the surface a client learns the key from
instructions	DarlingMcpInstructions.cs ×5 (get_query_store_top, get_query_trend, the trio), McpInstructions.cs ×4	say window_truncated, and where the row explains the block, why it renamed
web catalog	DarlingWebEndpoints.cs ×5	one clause each
prose	DarlingTrendReader.cs, DurationTrendRouting.cs, ViewerDataService.QueryTrends.cs, ViewerServerTab.QueryTrends.cs	the doc comments that named the wire key name the new one; the C# members keep their names (DurationTrendResult.Truncated, QueryHistoryResult.Truncated, TruncationSlack, DescribeCoverage — ViewerTrendRoutingPortTests pins the alias signatures byte-for-byte and they are untouched)

get_query_store_top on Lite publishes no window floor at all today (no effective_start, no flag) — nothing to rename there; its description stays as it was, and the two SKUs' descriptions for that one tool now differ by the Darling clause because the facts differ. The brief's premise that #3706 made the trend descriptions byte-identical across SKUs is not what the tree says — they already differ by the tier sentence ("On the per-collection (raw) route …" vs "Lite has one tier …") and no parity pin holds them equal; the shared clause is identical by construction on the nine that carry it.

The census, which is the part that keeps this renamed

McpPayloadContractCensusTests (Darling.Tests, sweeping both SKUs' tool sources):

• window_truncated joins SecondBoundCutKeys with the three files that emit it, and the class summary's residue sentence ("truncated carries two facts …") is gone. CutNoteKeys' truncation_note entry says which flag it sits beside on each tool.
• The payload-key matcher now reads the ordered-envelope idiom (envelope["key"] = …) as well as initializers. Both disclosure writers use it, so while the census matched initializers only, the window floor was invisible to it — which is how the homonym survived a roster that read as complete. The only cut-ish key ever written that way is this one (verified by enumerating every ["…"] = write in both tool directories), so the widening moves no other roster.
• TheWindowFloor_IsSpelledWindowTruncated_BesideItsReach_AndNeverBareTruncated_OnBothSkus is the fact the roster alone cannot state. A bare truncated is the page dialect and is permitted everywhere; what makes a truncated the homonym is its neighbour. So the population is every block on either SKU that writes effective_hours_back as a payload key — the innermost brace pair around it, walked over CSharpSourceWalker-strings-blanked code so a brace in a message cannot open or close it — and inside each, window_truncated and effective_start must be written and truncated must not. Offenders fail by file and line. The publisher set is an equality on (file, idiom, count): DarlingMcpDataTools.cs initializer ×1, DarlingMcpTrendTools.cs initializer ×1 + envelope ×1, McpQueryTools.cs envelope ×1.
• The discriminator is the reach vocabulary, not window_start as the brief proposed: get_query_heatmap publishes window_start / window_end (the requested axis) beside a bare truncated that IS a page cut — its cell cap, observed off limit + 1, with first_time_bin / last_time_bin saying which bins survived. A window_start rule would flag that tool for spelling its page cut correctly.
• TheWindowFloorChecker_FlagsTheHomonym_AndPassesAPageCut runs the checker on a synthetic offender (line-numbered), a synthetic page cut with a brace inside its message, and a synthetic envelope with a nested block — so a matcher that stopped matching reads as red, not as a clean sweep. Also executed live on the real tree: putting envelope["truncated"] back in Lite's WriteDisclosure reds the fact at McpQueryTools.cs:810 (envelope) and reds the roster equality (emitted on [DarlingMcpDataTools.cs, DarlingMcpTrendTools.cs]).
• EveryWindowFloorTool_CarriesTheSharedClause_AndNoOtherToolDoes: the nine tools whose attributes carry McpHelpers.WindowTruncatedDescription, as an equality; the constant executed for the key, the reach, "not a page cut", and the wire-change sentence; and by reflection the Darling five publish it in the attribute a client reads.

Pins that read the key on both SKUs moved with it: DarlingPerformanceTrendsReadTests (5 asserts + the six-key order block), QueryStoreTrendRoutingLiveTests (3), QueryStoreTopWindowTests (now reads the tool's own body — the old whole-file Contains("truncated") would have passed on any of the file's page-cut neighbours), Lite PerformanceTrendsToolTests (4 + block) and TrendEmptyParityToolTests (4). New: ServerPageTabsTests.TheQueryTrendDrillDown_ReadsTheWindowFloorOffItsOwnKey.

What this PR does NOT do

• Lite's get_query_store_top gains no window floor. It never had one (no effective_start, no flag); adding the reach probe is a feature, not a rename. Noted for the item.
• deprecated/Dashboard is untouched and unaffected: its McpQueryTools.cs publishes neither effective_start nor the flag, and it is not an MCP client, so it neither emits nor reads the key.
• truncation_note on get_query_store_top keeps its name. It is the prose beside the flag, the flag is now window_truncated, and the roster says so.
• The C# members keep their names — the rename is the wire key. ViewerTrendRoutingPortTests pins DescribeCoverage's (DateTime EffectiveStartUtc, bool Truncated) tuple and TruncationSlack's alias byte-for-byte; both stand.
• Item 17's second clause — the two stored-interval spellings — is inventoried, not renamed. Table below.
• No schema rung, no PgTarget*, no CHANGELOG.md.

Inventory: sample_interval_seconds vs interval_seconds on the wire (item 17, second clause — NOT renamed)

The brief's whole-tree counts (git grep -c '"interval_seconds"' = 19, '"sample_interval_seconds"' = 70 at this head) are dominated by SQL column names, collector projections and test pins; the wire sites — payload keys in the two MCP tool directories — are twelve, on six tools:

SKU	tool	file	key	what the value is
Darling	get_latch_stats (row + latest block)	DarlingMcpLatchSpinlockTools.cs ×2	interval_seconds	reader's LatestIntervalSeconds, null when unknowable (restart row)
Darling	get_spinlock_stats	DarlingMcpLatchSpinlockTools.cs ×1	interval_seconds	same, null when unknowable
Darling	get_pg_io_trend	DarlingMcpPgTrendTools.cs ×1	interval_seconds	derived span between DISTINCT collection times (not a stored column)
Darling	get_file_io_stats	DarlingMcpDataTools.cs ×1	sample_interval_seconds	the raw stored column as collected (0 on a restart row is passed through)
Darling	get_perfmon_trend	DarlingMcpTrendTools.cs ×1	sample_interval_seconds	raw stored column
Darling	get_resource_semaphore	DarlingMcpMemoryGrantTools.cs ×1	sample_interval_seconds	is > 0 ? … : null — the null-when-unknowable fact under the raw spelling
Lite	get_latch_stats, get_spinlock_stats	McpLatchSpinlockTools.cs ×2	interval_seconds	KnownInterval(...), null when unknowable
Lite	get_file_io_stats	McpIoTools.cs ×1	sample_interval_seconds	raw stored column
Lite	get_perfmon_trend	McpPerfmonTools.cs ×1	sample_interval_seconds	raw stored column
Lite	get_resource_semaphore	McpMemoryTools.cs ×1	sample_interval_seconds	is > 0 ? … : null — null-when-unknowable under the raw spelling

Read at each site, the split is mostly clean — interval_seconds = "the seconds this rate was divided by, null when there were none", sample_interval_seconds = "the column as stored" — with one crossed pair: get_resource_semaphore on both SKUs publishes the null-when-unknowable fact under the raw spelling. The #3703-style pass would be: that pair to interval_seconds (both SKUs, one consumer sweep), and a description clause on the two raw-column tools saying 0 means the collector could not difference. The web client reads neither key (zero hits under wwwroot/js). Not done here — it is its own lane.

Executed on this Mac

Builds -c Release -p:EnableWindowsTargeting=true, 0 warnings: PerformanceMonitor.Common, Darling.Storage, Darling.Service, Darling.Viewer, PerformanceMonitorLite, Darling.Tests, Lite.Tests. node --check on server-tabs.js.

Darling.Tests through a Darling.Tests-named net10.0 harness against a throwaway timescale/timescaledb:2.28.1-pg18 (context desktop-linux, torn down): the six named censuses (RepoFileAdoptionTests, StragglerCommandTimeoutTests, LiveCleanupConversionRatchetTests, MeasurementContractCensusTests, McpPageContractTests, FactSourceRegistryTests) plus FleetIdentifierScrubTests, PgTargetMcpSurfaceTests, McpLatestSnapshotStampTests, DocCommentHygieneTests, McpPayloadContractCensusTests (67), QueryStoreTopWindowTests, ServerPageTabsTests, DarlingMcpTrendToolsTests, BaselineDiscontinuityRenderTests, DarlingWebEndpointsTests, McpZeroIsAMeasurementTests, DarlingMcpDataToolsTests — green; live: DarlingPerformanceTrendsReadTests (raw + empty envelopes), DarlingPerformanceTrendsTierRoutingLiveTests (hourly route, the unserved-head empty envelope), QueryStoreTrendRoutingLiveTests (rollup+raw, below-floor, tail-only), DarlingMcpDataToolsLivePostgresTests (get_query_store_top's envelope) — 6/6 green reading window_truncated off every route. Whole harness: 10,969 tests, 164 red = the platform baseline (159 on lane-3778's dev run) + 3 live fixture-path artefacts of the harness (Fixtures/SystemHealth/*.xml not copied) + the two #3787 dev reds #3790 has since fixed (this branch is rebased onto it). Lite through a Lite.Tests-named harness: PerformanceTrendsToolTests, TrendEmptyParityToolTests, McpMissMessageParityPinTests, McpDescriptionTruthPinTests, BaselineDiscontinuityRenderCensusTests, CrossAppMcpToolInventoryPinTests, MeasurementContractCensusTests, McpLatestSnapshotStampTests green; McpPageContractTests 19/21 with the two PresentationFramework wait-stats reds reproduced on dev by #3702 / #3703 / #3706.

First executed in CI: nothing on this surface — every changed assertion ran here.

Partial #3653 — item 17 (window_truncated)