Complete the Darling README's PostgreSQL collector inventory (27 of 27), pin it, and retire the last seven-collector-era prose

[erikdarlingdata]

Fixes #3261 (close on merge)

Stacked on #3260 (fix/3247-runbook-staleness): this branch contains #3260's two commits because both edit Darling/README.md's PostgreSQL section and the "What gets collected" work builds directly on #3260's corrected rows. Merge #3260 first and this PR's diff collapses to its own two commits (README + tests/comments/tool description). No CHANGELOG.md, per the wave protocol — the entry was reported for the consolidated splice.

What #3261 listed, and what was done

1. "What gets collected" covers 12 of 27 collectors — the table now carries all 27. Each new row is derived, not invented: source and behavior from the collector's own doc comment and its migration rung's header (PgMigrations.cs V96–V106), cadence/retention verbatim from CollectorScheduleDefaults.cs, gates from AppliesTo/RunsPerDatabase. Extension-miss degradation is worded "a non-fatal skip naming the missing install" so the rows stay true if the in-flight EXTENSION_MISSING reclassification (fix-3240's lane) lands in either order.
   • Pinned: new ReadmeDerivedCountPinTests.WhatGetsCollectedTable_NamesEveryPostgresCollector requires the table's row-leading names set-equal to the catalog's PostgreSQL roster in both directions. It is a bespoke fact rather than a SetPins entry because collector names legitimately appear inside other rows' source cells ahead of their own row (pg_replication_slots is a source of pg_xmin_horizon), which defeats the shared drop-one-token mutation proof; it carries its own drift proof that deletes a whole row instead. Proven red externally too: deleting the pg_wraparound_stats row failed the fact naming that collector, then the file was restored.
2. V1 ladder row's "twenty PostgreSQL tables" — now "27", with the exact rung list (V63–V69, V71, V83–V94, V96–V99, V102, V103, V106), and the notable-rungs table gains the seven rungs it never listed: V96 wait sampling, V97 kernel stats, V98 predicate stats (all PostgreSQL wait/CPU/predicate extensions: buildable locally, dark on Aurora #2603), V99 plan capture (Store PostgreSQL plans: query_plan_dim is built for SQL Server plan XML, and PostgreSQL plans are JSON #2566/PostgreSQL has no execution plan capture, which is where a DBM comparison is lost #2538), V102 server configuration (PostgreSQL configuration is never collected: no pg_settings read, and no way to see what changed #2658), V103/V104 deadlocks + identity index (PostgreSQL deadlocks are a number with no detail, and the log already carries the whole graph #2661), V106 instance CPU (Postgres High CPU alert isn't a light port — there's no instance-level CPU collector to alert on #2719). Every row's content comes from that migration's own header comment.
3. Stale "seven collectors" code comments — verified each before touching, per the issue's own instruction:
   • DarlingCollectorRunner.cs:1099 ("Six of the seven ... take that path") — present tense, now false (7 of 27 fan out today): era-anchored to "then shipping ... took".
   • DarlingCollectorRunner.cs CreateTargetConnection remarks and CollectorRunnerConnectionEngineTests header — same two-word era anchor.
   • CollectorRunnerConnectionEngineTests assertion message ("is what made six of seven ... fail") — already past-tense history, left alone.
   • PgMigrations.cs:3138 / PostgresMajorVersionRegistryTests "Seven PostgreSQL collectors branch on postgresMajorVersion" — left alone, verified CURRENT: exactly seven collectors pass the major version into BuildQueryText today (autovacuum, index usage, io, replication slots, session states, statement stats, write stats). This is a different seven than the collector-count seven; the issue flagged it and verification cleared it.
4. get_pg_top_queries MCP description — no longer opens "for Amazon Aurora PostgreSQL targets". It now states the Self-hosted PostgreSQL has no top-queries answer at all: pg_statement_stats is Aurora-only #2625 truth: works wherever pg_stat_statements is installed, vanilla view elsewhere with Aurora-only columns null; the I/O source breakdown sentence is scoped "On Aurora", and peak memory is marked "(Aurora only)" — wal_bytes stays unscoped because the WAL columns are core pg_stat_statements, not part of the six Aurora-only columns. No test pins the old opener (grepped).
5. "What it does not do yet" / tab counts — re-verified against dev as the issue asked: "seven inner tabs" and the tables-no-charts claims hold; no edit needed.

Also updated in the same stroke: #3260's "the Performance Insights CPU read that is not in the table above" clause — true when written, false once the table gained the pg_cpu_utilization row — trimmed to just the epithet.

Collision lanes

• Name the pg_read_file grant pair on the self-hosted log readers' PERMISSIONS hint #3251 (grant regions): untouched — this PR edits README lines ≥685 and the ladder; Name the pg_read_file grant pair on the self-hosted log readers' PERMISSIONS hint #3251's README hunk sits at ~676 with clear separation.
• fix-3240 / EXTENSION_MISSING: no status-name claims added; the one degradation mention is status-neutral.

Verification

• ReadmeDerivedCountPinTests (all existing pins plus the new fact), CollectorRunnerConnectionEngineTests, PostgresMajorVersionRegistryTests, CrossAppMcpToolInventoryPinTests, PerCycleTimingLogLevelTests — 18/18 pass via the MTP runner; new-test red proven with an injected row deletion.
• Build: 0 warnings in the touched projects.
• README byte census clean: pure CRLF, zero lone LF/CR; 27 row-leading pg_* names counted.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AtjWJJ3NnSkDSrPBsGywoZ

[claude]

Reviewed. This is a documentation/comment/test-only PR — no T-SQL, no functional code paths changed (only doc comments in DarlingCollectorRunner.cs/CollectorRunnerConnectionEngineTests.cs, the get_pg_top_queries MCP tool description string, Darling/README.md, and a new pinned test).

Checked and confirmed accurate:

• New test (WhatGetsCollectedTable_NamesEveryPostgresCollector): the leading-name regex correctly handles the arrow-notation rows (pg_replication_slots` → `collect.pg_replication_slot_stats`, pg_blocking` → `collect.pg_blocking_edges`) because PgReplicationSlotsCollector.Name / PgBlockingCollector.Name are "pg_replication_slots" / "pg_blocking" — matching the row's leading cell, not the arrow target. The bespoke whole-row drift proof (vs. the shared token-drop mutation) is justified given pg_replication_slots legitimately appears inside pg_xmin_horizon's row before its own.
• README additions: the V96–V99/V102/V103/V106 rung descriptions match PgMigrations.cs's own header comments, and the V1 row's "27" ties out against the rung list (7+1+12+4+1+1+1=27).
• get_pg_top_queries description update: matches the actual nullable behavior in DarlingMcpPgStatementTools.cs (Aurora-only columns null-not-zero, peak memory Aurora-only).
• No Lite/Darling parity issue — PostgreSQL collectors and this README are Darling-only; Lite has no PostgreSQL surface.
• CHANGELOG.md omission is explained in the PR body (wave protocol, reported for a consolidated splice) and matches repo convention.

No correctness, security, or performance concerns. Nothing to flag inline.