Changelog: third splice — 55 entries from the #3653/#3691 waves, the Azure quad, the night queue, V137 and the post-crash repairs (66 merged PRs, 11 entries reconstructed from PR bodies)

[erikdarlingdata]

What this is

Sixty-six pull requests merged to dev since #3751 (the second splice, 2026-09-19T15:51Z), and by the repository's own rule none of them touched this file — each returned its entry to a buffer instead, so that the changelog is written once rather than merged sixty-six times. This is that one splice: fifty-five entries, one file, text only, and nothing outside ## [Unreleased] and the link-definition block moves.

The waves it covers, in the order they merged:

• The Azure quad ([BUG] long_query_completions XE session cannot be created on Azure SQL DB: 'sqlserver.server_principal_name' is not an available action #3753, [BUG] database_scoped_config collects nothing on Azure SQL DB: three-part [db].sys.sp_executesql is not supported there #3755, [BUG] query_store_health collects nothing on Azure SQL DB logical-server registrations: the same three-part [db].sys.sp_executesql that #3755 removed from database_scoped_config #3764, [BUG] Azure collector failures are swallowed into status=SUCCESS, and get_collection_health then reports HEALTHY with a finding that says the opposite of what happened #3754) — three collectors that were dead or silent on Azure SQL DB by construction, and the swallow that let the fourth report SUCCESS over it.
• The night queue (MCP refusals have no status word: ~400 validation bails return a bare sentence, and seven PostgreSQL refusals now answer HTTP 500 on the web where they answered 200 before #3719 (Q11 residue) #3739 through The materialization hole scan is silent when it finds nothing — a zero-hole start and a start where the scan never ran leave the same log #3756, The CPU collector's watermark is the target's LOCAL sample_time and its dedup drops every sample at or below it, so the autumn fall-back hour never reaches the store — the other half of #3744's blind hour #3778, The bundled store authors its memory knobs from the box but leaves max_wal_size at PostgreSQL's 1 GB default — derive it from disk headroom (≤16 GB) in the same versioned, heal-on-start block #3802, Restore missing_indexes[].create_statement on the plan tools — #3696 dropped it citing a 'no-missing-index-recs rule' that was never made; DDL recommendations are legitimate product output (RCSI precedent) #3805, pg_statement_stats dead on 23/50 PostgreSQL clusters since .453: the pg_stat_statements_info read fails the whole collector where the extension is below 1.9 or outside public — gate on to_regclass, not the server major; fix the false EXTENSION_MISSING text #3818, Held retention policies only arm on service restart - re-evaluate them on a periodic cadence too #3812) — the refusal envelope's status word, the CPU gate's UTC identity and the collector watermark behind it, the CONFIG_CHANGED anchor, two pair gates, the long-running-query exclusion order on both SKUs, the hole scan's unconditional line, the headless collector switch, the managed store's derived WAL ceiling, the restored CREATE INDEX text, the statements-epoch gate, and the retention re-arm.
• The Brains-review campaign: deferred structural residue (from #3538 / #3539 / #3540 / #3541) #3653 brains-review residue — Rule 6's C# census, the CONFIG_CHANGED A10 second slice, the rendered discontinuity marker, the daily calendar's NULL-not-0, the window_truncated homonym, and item 17's remaining pins.
• The PostgreSQL-target analysis engine: v2/v3 slices and the v1 lanes' residue (from #3542) #3691 PostgreSQL-target engine, waves 3 through 6 — the target clock, the autovacuum-disabled card, sampled waits with their own baseline, collection caveats, the saturation numerator, v3 plumbing, V136's size series and host memory, buffer composition, plan regression, kernel CPU, sequential-scan evidence, the per-key baseline seam, memory composition, the bad actor's own normal, database growth, and two absorb-what-the-lanes-reported passes.
• V137 and its consumers — the rung itself, then the capture-mode publish, the store TOAST and checkpointer surface, the route knob's store home, and the clutter view's two viewers.
• The post-crash repairs — the dev-red arithmetic, the companion-object remedy, the collector-regression band, and the TimescaleDB availability re-probe.

Forty-four entries verbatim, eleven reconstructed

Forty-four bullets are verbatim from the two buffers (trailing whitespace stripped, nothing else). Those buffers were casualties of the 2026-09-20 reboot — /tmp went with it — and were recovered from session transcripts; they now live on the bus at design/pending-entries/ rather than under /tmp, which is the lesson that wipe taught.

Fourteen merged pull requests had no buffer entry at all, because their lanes died in the crash or the merging session wrote none. Rather than leave a schema rung and four shipped surfaces undocumented, those are written here from the pull requests' own bodies in this file's voice, and each is named so a reader knows which prose is reconstructed rather than a lane's own words:

Entry	Section	Source
Darling V137 / Lite v64: eight nullable columns	Added	#3813 (its buffer header was a placeholder waiting on a report that never came)
Nothing could say which databases were cluttering Query Store	Added	#3814 (same)
get_query_store_health skipped the one knob that names a plan-churn factory	Added	#3823
The store could say how big its plan dimension's TOAST file was and not how full	Added	#3826
The uncorroborated-finding route knob said FILE-LEVEL	Added	#3828
The Query Store clutter view had no viewer	Added	#3841
Four census pins the V137 wave outran	Fixed	#3806, #3821, #3822, #3825 folded into one ref-less bullet
Two counts nothing re-derived turned dev red on three checks	Fixed	#3837
The companion-object remedy told twenty-three clusters to UPDATE an extension none of them had	Fixed	#3839
A collector that STOPPED producing read as one that never produced	Fixed	#3840
TimescaleDB availability was decided once, by a probe whose failures last a minute	Fixed	#3842

The four census pins are folded rather than given four bullets because they changed no product file and nothing a reader of the product sees — but they are not omitted, because the shape they record is worth keeping: a "no reader names this column yet" pin has one honest lifetime, from the rung's merge to the first consumer's, and here that was forty minutes.

Classification

Twenty-one entries append to ### Added (new capabilities, fact families, tools, columns and rungs), eight to ### Changed — a designed behaviour that now behaves differently: two pair gates (#3741, #3691's stock sampled arm), the CONFIG_CHANGED anchor moving to the trace line (#3740), the refusal envelope's wire shape (#3739), the WAL ceiling's derivation from disk headroom (#3802), a story's root naming its consumed leaves (#3691), the second fleet read's lineage stamps (#3691) and the bad actor's own-normal grading (#3691) — and twenty-six to ### Fixed. Order within each section is ascending pull-request number, as both prior splices did.

Two entries have no issue to cite and cite none, matching the file's ninety-six other ref-less entries. Every other citation is the issue the pull request closes or partially addresses, resolved against the API rather than inferred from a branch name; twenty-three of them needed a definition, all appended to the end of the [Unreleased] run so that run stays one ascending sequence (3517..3830), never the file's tail, which belongs to an older version.

What was checked

• python3 tools/changelog/changelog_archive.py verify — all PASS. No archive is touched and the split was not run.
• Every ([#N]) resolves to a definition in this file; no reference is undefined in either direction; no definition is duplicated.
• CRLF throughout, read and written as bytes — 3,494 CRLF, zero bare LF. The file is 469,093 bytes, well under the render ceiling.
• The diff is seventy-eight pure insertions in four hunks: twenty-one, eight and twenty-six bullets appended at the end of their own sections, and twenty-three definitions. Nothing else in the file is touched, and no other file is in the commit.
• ChangelogIndexAndArchiveTests executed locally, 5/5. A changelog-only change takes CI's docs fast path — build passes in about fifteen seconds with every compile and test step skipped — so that class does not run there. It was compiled into a one-file net10.0 xunit harness whose output directory sits inside this worktree, so the test's walk-up finds this tree's PerformanceMonitor.sln and reads this tree's CHANGELOG.md. Mutation control: deleting one definition this splice introduces reds EveryFile_ResolvesItsOwnReferencesAndRendersOnGitHub by name and by number, and restoring it returns the diff byte-identical.
• Client hygiene: the whole [Unreleased] section was scanned for the three forbidden name families and nine structural shapes — managed-database and cloud provider hostnames, instance identifiers, account identifiers, resource names, network addresses, engine instance paths and network share paths. Zero hits. Every population statement in the reconstructed prose is a count with no name attached, in the register the file already uses.
• Every buffer header marked [armed] was confirmed MERGED with its merge commit before its bullet was taken, and so was every other pull request in the two buffers — fifty-two of fifty-two.

What it does not do

No product file, no test file, no archive, no census. Auto-merge is deliberately not armed.