A15/A16 remainder: the six named drifts were already landed (#3665/#3696/#3697) — this pins the three that had no pin: get_ag_health names its reader's traps, the catalogue's duration-trend line says rates not percentiles, server-tabs.js reads the Stamped tools by captured_at (#3653)

[erikdarlingdata]

The lie

Item 17 of #3653 (A15/A16, description drift + envelope vocabulary) still listed six drifts as open after #3793: audit_config claiming edition-awareness and answering no_config_data for a PostgreSQL target; plan top_operators silently Take(10) and switching its ranking basis; missing_indexes.impact an unlabelled statement-scoped percentage; get_ag_health hiding its reader's two documented traps; the web catalogue calling the query duration trend "percentiles"; and server-tabs.js reading the four StampedUnderCollectionTime tools by the old key. Read at source on dev 32096b62, all six were already delivered — by #3665 (the not_collected envelope), #3696 (the other five descriptions and the plan formatter), and #3697 (the JS tile). The item text was stale, not the work.

What was NOT delivered was the pin. Three of the six fixed texts had nothing holding them: the AG description's two trap sentences, the catalogue line's "not percentiles", and the Sessions tile's captured_at read. Each is a fix that fails nothing when it regresses — a description tidy drops a sentence and the reader's doc comment goes on being right in a file no agent opens; a catalogue edit reintroduces a word no payload backs; a JS key drifts back and the tile renders a dash where the snapshot's age was, still drawing, still "working". This PR is those three pins, and nothing else changes on the wire.

Mechanism, per pin

DarlingMcpAgToolsSurfaceTests.Description_SurfacesTheReadersTwoDocumentedTraps. Anchored first on DarlingAgReader's own bolded paragraph headers (<b>Latest collection per server.</b>, <b>What the badge deliberately does NOT band: lag and queue depth.</b>), so the pin fails with "re-anchor" if the reader stops documenting a trap rather than passing against prose that no longer exists. Then the description, for each trap's mechanism, consequence and tell: the collectors write NO row for a server with no AGs → last non-empty snapshot → collection_time; lag and queue depth are NOT banded → secondary_lag_seconds → is_suspended and the DMV's 0 lag while data movement is suspended. Then the instructions' AG paragraph, which carries the same two in its own words beside the perspective trap (LOCAL-replica-only columns) it always had.

DarlingMcpTrendToolsSurfaceAndSqlTests.QueryDurationTrend_CatalogueLine_NamesTheRatesThePayloadCarries_NotPercentiles. Both ends. The catalogue line (DarlingWebEndpoints.CatalogDescriptors["get_query_duration_trend"]) must name ms/sec and executions/sec and say not percentiles; SerializeTrend — the helper the duration trio shares — must still project elapsed_ms_per_second = p.Value and executions_per_second = p.ExecutionsPerSecond and contain no "percentile"; and across the whole catalogue the word may appear only negated ((?<!not )percentiles? matches nothing), because no read here computes one. A read that genuinely returned a percentile would add itself beside that sentence rather than fall through it.

ServerPageTabsTests.TheStampedReads_AreReadByTheirStamp_NeverTheRetiredKey. Derived, not listed. The roster is McpLatestSnapshotStampTests.LatestTools filtered to Stamped and SearchBound — every read of those shapes publishes its stamp once at the top as captured_at and nothing under collection_time anywhere in its payload (the census's negative sweep refuses the top-level key, and none of those fifteen tools projects the word into a row either), so any descriptor the page binds to one of them that names key: "collection_time" is a dead read, tile or column alike. The page side is parsed by bracket balance: every stat(…), table(…) and fanout(…) call, its first read-shaped literal, and every *_STATS / *_COLUMNS token in it, resolved to the module's top-level const NAME = [ … ]; arrays. Twelve bindings over the roster are examined today; the four reads the item named must each be reached or the test fails on the parser rather than passing on nothing. Windowed reads are excluded on purpose: get_memory_grants / get_resource_semaphore rows carry a per-row collection_time that IS a series column. The positive half is the one tile the item named — SESSION_STATS must show { key: "captured_at", label: "Collected", format: "reltime" } — so the pin cannot pass by the tile being deleted.

Why this shape

Pins only, no prose edits, because every sentence the item asked for was already on the surface and re-wording a correct description to "own" it would be churn on the wire an agent reads. The AG pin is phrase-matched rather than paraphrase-tested because these are the exact words a caller searches a description for. The JS pin is derived from the stamp roster rather than naming the four reads, so a fifth read joining the roster is covered the day it joins — the four are asserted as a reach check on the parser, not as the population.

What this does NOT do

• Does not rename get_resource_semaphore's interval key. 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) #3793 inventoried the crossed pair: get_resource_semaphore (both SKUs) publishes the null-when-unknowable interval under the raw spelling sample_interval_seconds beside interval_known, while get_file_io_stats / get_perfmon_trend publish the raw stored column under the same key with 0 meaning unknowable, and get_latch_stats / get_spinlock_stats publish null-when-unknowable under interval_seconds. Every one of those is described truthfully at its site today (the two raw-column descriptions already say what a stored 0 means). One fact, two encodings, three spellings — unifying it is a wire change on a well-described surface, which is the maintainer's ruling, not this lane's. Stated on the issue.
• Does not restore the missing-index DDL. The brief for this lane said "do NOT remove the DDL; note the collision as a ruling" — mute_analysis_finding stops writing the hash as the path and stops registering the same mute twice, remove_server can remove a server that never connected, and eight MCP descriptions stop saying what the code does not do (partial #3653 A15/A16) #3696 had already removed the paste-ready CREATE INDEX from the MCP missing_indexes[] rows (column lists + impact_basis + a hint-not-design note stay; the plan viewer still renders the DMV text for a human). The ruling now runs the other way: whether the MCP surface stays DDL-free. Left as it landed; stated on the issue.
• No PgTarget*.cs, PgBaselineProvider.cs, DarlingPgSessionStatesReader.cs. No schema rung. No CHANGELOG.md. No Lite twin: Lite hosts no AG tool, no web catalogue and no server-tabs.js, so there is nothing on that SKU for these three pins to hold.

Verified here

Builds -c Release -p:EnableWindowsTargeting=true, 0 warnings: PerformanceMonitor.Darling.Viewer, PerformanceMonitor.Darling.Service, Darling.Tests.

Darling.Tests through a Darling.Tests-named net10.0 harness on this Mac (the whole test tree compiled against the built dlls):

• The three new pins red against their defects, then green: the AG description with the lag and queue depth are NOT banded sentence removed → red on that phrase; the catalogue line reverted to Query duration percentiles over time. → red on ms/sec and again on the whole-catalogue sweep naming get_query_duration_trend; the SESSION_STATS tile keyed back to collection_time → red naming get_session_stats/SESSION_STATS and the retired key. All three green on the head.
• Censuses green: RepoFileAdoptionTests, McpPageContractTests, McpPayloadContractCensusTests, MeasurementContractCensusTests, FactSourceRegistryTests, DocCommentHygieneTests, McpLatestSnapshotStampTests; the three host classes in full (ServerPageTabsTests, DarlingMcpAgToolsSurfaceTests, DarlingMcpTrendToolsSurfaceAndSqlTests).

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

Partial #3653 — item 17 (A15/A16)