Record the CHANGELOG entries for eighteen changes across eleven merged pull requests, and strike an unsound claim from #3199's

[erikdarlingdata]

Applies the collected [Unreleased] entries for the pull requests merged to dev after #3213, and strikes an arithmetic claim that pass shipped which #3226 has since shown to be unsound. Lanes do not edit CHANGELOG.md -- every one of them appends to the same [Unreleased] block, so a per-PR edit conflicts with whichever sibling merges first -- and the entry text rides each pull request's own description instead.

What is in the diff

CHANGELOG.md only, 35 insertions and 1 deletion: 18 entries prepended inside ## [Unreleased] -- 6 under ### Added, 12 under ### Fixed, none under ### Changed -- plus the 16 link-reference definitions those entries need, plus the one edited line described below. Entries are ordered newest-merge-first within each section, matching the file's existing order and asserted against each pull request's own mergedAt. Nothing else already in the block was edited, reworded or reordered.

15 definitions are appended in ascending order at the foot of the file. [#3189] is below [#3196] and joins the same run at its own ascending position, between [#3188] and [#3192], so all 15 definition runs keep exactly the ordering they had -- 9 of 15 ascending, before and after.

The range, recomputed rather than inherited

Eleven pull requests merged to dev after #3213's own merge at 2026-09-09T14:29:33Z. Ten were there when this was cut, counted three ways that fail differently -- gh pr list --base dev --state merged filtered on mergedAt, the search API on is:merged base:dev merged:>=, and git log --first-parent over the merge commits, all returning the same set: #3212, #3216, #3218, #3220, #3223, #3224, #3225, #3226, #3227, #3231. #3230 then merged at 19:31:55Z while this was being verified and is absorbed, which makes it the newest merge in the range and the first entry under ### Added.

The three-way count is not ceremony here. gh pr list returned exactly 200 records against a --limit 200, so that enumeration was truncated at the boundary and sorted by creation date -- the shape that silently drops an old pull request merged late. The search API's total_count and the first-parent walk are what make the set trustworthy rather than the list command.

The boundary is checked from the other side too: the highest citation anywhere in [Unreleased] is #3200, and none of the eleven pull requests, nor any of the 17 numbers this diff cites, was already cited or already defined. Nothing merged inside the range is uncovered, and nothing is covered twice.

#3213 itself earns no entry, on the precedent it recorded: no batch pass is cited anywhere in the file. Nothing else is open against dev.

Where the text came from

Seven of the ten carry their own entry text under a ## CHANGELOG … heading in the body, and it is used as written: #3216, #3218, #3220, #3224, #3225, #3226 and #3227. Shaping is limited to what the block's existing entries require -- bolding the opening clause where a lane wrote a plain bullet (#3220, #3226, #3227), moving a trailing (#NNNN) into the ([#NNNN]) position after the bold lead, and terminating a sentence that ends mid-thought (#3218 and #3224 both do; a period is added and nothing else).

Three carry none: #3212, #3223 and #3231, and #3230's arrived late -- see below. The queue directories were searched rather than assumed empty, since #3213's pass had to be corrected for missing a handoff file and #3150 records that 11 of its 66 entries came from one. Neither inbox/ nor outbox/ holds entry text for any of them -- #3223's own lane report says only "No CHANGELOG.md edit", and the files that mention #3212 are audits of it. Those four entries are therefore written here from the pull request descriptions, which is stated rather than left to look like lane prose. Measured rather than asserted: normalising away the shaping and matching every entry against the long bullets in the bodies that do carry text, 14 of 18 score >= 0.995 against their source and exactly 4 score < 0.04 -- and those 4 are precisely the ones above.

#3230's entry is the lane's own block, and the first version of it was not. When its description was read at 19:43 it carried no entry-ready block, so the entry was written from the description like the other three. The lane added a ## CHANGELOG entry text block at 20:00:22Z, and that block replaces the entry wholesale rather than the entry being patched -- it names only what shipped. What the description did lead with was a measurement section taken at c9f04f3fe, before #3227's rebase moved ViewerDataService.Deadlock.cs LastTranStartedLocal onto FormatServerClock, which took that member out of both the inventory and its string-literal subset. KnownTruncatedRanges ships 30 entries -- verified by counting the array two ways that agree, at TsqlConventionGuardTests.cs:1414, closing at 1447, with LastTranStartedLocal absent -- while the retired entry said 31, and carried 13, 544, 510, 2,238 and 42,927 besides. The 544 and 510 predate the content-trim and cannot be recomputed from shipped code at all; the file and declaration totals move with every commit, and a release note should not carry a number that drifts.

The staleness is in the merged code too, not only in the description. KnownTruncatedRanges' own doc comment says "Editing 31 member bodies", "the current 31" and "Thirteen of these strand a string literal", against an array of 30 -- so both sources a lane could have read agree on the wrong number. That is a counted claim welded to a numeral, in a file whose entire subject is instruments that report something other than what they check. Reported rather than touched here, since this diff is CHANGELOG.md only.

The entry's one surviving quantity is cross-checked against the shipped array by the verification battery, so it cannot restate a count the code does not have. That check is what this class needed: the retired figures were all internally consistent with each other and with the prose they came from, so nothing short of reading the array would have caught them.

Four pull requests contributed more than one entry because they did more than one thing: #3226 five (its own five staged bullets), #3227 three (two Fixed, one Added, as its lane assigned them), and #3223 two, one per guard -- they are two different suites with two different defects, and merging them would have buried the second.

Two citation judgements

Citations are issue numbers where an issue exists and pull-request numbers where one does not, which is what the file already does. Each of the 17 was resolved against the API's own pull_request field rather than guessed: 12 are issues and 5 are pull requests (#3223, #3224, #3226, #3230, #3231), and the 5 use the modern /pull/N form -- /issues/N for a pull request is a third of the file's population but the split is by age, and the highest such is #2928.

• Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226's fourth and fifth bullets carry no citation of their own. The fourth (CONTRIBUTING.md paths, three deleted .cmd scripts) has no issue behind it and cites the pull request, [#3226]. The fifth is the dependabot.yml bump-review note, and an issue does exist for its subject -- Azure.Identity is unpinned and transitive, so #3218's broker-free guarantee rests on a dependency-graph fact nothing tracks #3219, "Azure.Identity is unpinned and transitive" -- so it cites that, which is the file's rule rather than my preference. Azure.Identity is unpinned and transitive, so #3218's broker-free guarantee rests on a dependency-graph fact nothing tracks #3219 is consequently cited twice, since Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224's own entry also references it; that is precedented rather than a collision, as 32 numbers already lead more than one entry and 133 NpgsqlCommand sites still inherit Npgsql's undocumented 30s default timeout #2874 leads fifteen. [#3223] likewise leads both of its entries. No definition is duplicated.
• Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224's entry cites its own number, corrected in the second commit. As first pushed it led with [#3218], the pull request it is a follow-up to -- which is the lane's own wording, but in the citation position it resolves a reader to a different change. No issue exists behind Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224, so the rule above puts its own number there, with the follow-up relationship kept as prose. It was the only one of the ten with no issue behind it whose number went uncited, and that is now checked as a property rather than spot-read: for each of the eleven, either an issue exists and is cited in its place, or the pull-request number is cited, with no exceptions -- which is also what placed [#3230]. The [#3218] definition is swapped for [#3224] rather than added, since nothing else cited it -- so the count stays at 15 and the run stays ascending. Found by the review bot; the finding was verified independently before it was acted on.
• ### Added versus ### Fixed for the doc-shaped changes is decided on measured precedent, not taste. Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226's dependabot.yml note goes under Added because Pin the README's extension-dependent collector list to PgExtensionAvailabilityCollector's roster #3187 -- a README.md paragraph, "declared rather than described" -- already sits there. Scope the de-skew's accuracy claim where the column outlives a DST transition #3231 is doc-only and goes under Fixed on Document the PostgreSQL grant and extensions that decide whether a collector returns anything #3184's precedent, which Record the CHANGELOG entries for thirty-three changes across twenty-six merged pull requests #3213 recorded: the prose did not merely lack detail, it left an accuracy claim unscoped in the direction that fails silently, which is Scope the de-skew's accuracy claim where the column outlives a DST transition #3231's own account of itself.

The #3199 correction, which is the one deletion

#3213's pass shipped the #3199 entry containing a claim that PR #3226 removed the pin for, and #3226's lane filed a note about it precisely because CHANGELOG.md is not a lane's file to edit.

Before:

Raising the deadline was arithmetically closed rather than merely unattractive: the shared constant is pinned at 10 sites × N < 60 s, so N < 6, and a private constant for this one call caps at 14 s with nothing measured saying 14 is enough.

After:

A private constant for this one call caps at 14 s with nothing measured saying 14 is enough.

Both operands were wrong, not just the framing. DarlingWorker.SweepWatchdogSeconds -- the 60 -- is fed timestamps stamped at per-server collection-body launch and only selects a log level; it never observes the serial loop and cancels nothing. And the 10 is a census of the commands carrying SerialLoopSeconds rather than of the chain, while the same thread also awaits CollectionSweepSeconds, AlertPassCommandTimeoutSeconds, AnalysisCommandTimeoutSeconds and two 300 s budgets -- so 10 × 5 understated it. No replacement inequality is added, because #3226 established that nothing bounds that chain in aggregate; the entry it landed alongside in this same diff says so. The surviving clause is untouched and sufficient on its own, and the rest of that line is asserted byte-identical.

Verification

CI takes the docs fast path on a change that ships no code, so a green build in seconds has run nothing and is not the evidence here. What was checked, all of it re-parsed from the written file by a script that imports nothing from the assembler:

• git diff --numstat origin/dev reports 35 insertions and 1 deletion, and CHANGELOG.md is the only path in the diff, and the single removed line is asserted to be the Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199 entry. A CRLF renormalize -- which this file has suffered twice, in Fix CHANGELOG.md: restore CRLF after an accidental LF renormalize (#2694 fallout) #2696 and Fix CHANGELOG.md: restore CRLF after a second accidental LF renormalize (#2695 fallout) #2697 -- would show as thousands of both. The file is read and written as bytes, and CRLF integrity is asserted before and after: 3,591 CRLF, zero bare LF, no BOM.
• Blob line count 3,557 -> 3,591, a delta of exactly 34, which is the 34 genuinely new lines. .gitattributes normalizes the blob to LF, so LF there is expected and is not evidence of anything.
• Counted two ways that fail differently, and reconciled. By the - ** prefix: 19 lines, which is 18 new entries plus the re-added Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199. By a leading-([#N])-citation match, which would undercount any malformed lead: also 19. The second method is what caught an off-by-one in my own first reconciliation -- I had added the Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199 line twice, once as an entry and once as itself. Re-parsing the written file per section independently gives a delta of exactly 6 / 0 / 12.
• Citations against definitions in both directions, extracting every [#N] on a line rather than the first -- three of these entries cite two numbers and a first-match extractor would report false orphans. Every citation added resolves, every definition added is cited, no number is defined twice anywhere in the file's 1,247 definitions, and none of the 16 already existed. The only definitions cited by nothing remain Charts should show visible flat line at zero instead of empty #85 and Deadlock pipeline latency: too many sequential hops before data reaches charts #86, which pre-date this change, and the set of dangling citations is empty before and after.
• Structure unchanged: 23 ## [ headings and 81 ### headings, identical as ordered lists before and after, and [Unreleased]'s three subsections still Added, Changed, Fixed in that order.
• Definition runs: 15 before and after, the first 14 byte-identical, and the last run's pre-existing members in their original relative order with the new numbers interleaved at their ascending positions.
• Markdown well-formedness of all 19 touched entry lines: balanced backticks and **, a bold lead followed by a citation group, and a plain-hyphen separator on every one. All 16 definitions match ^\[#N\]: .../(issues|pull)/N$ with the number equal to the URL tail.
• The added text was scanned for RDS endpoints, ARNs, 12-digit account ids, monitor host names, dotted-quad and RFC1918 addresses: clean. The one dotted-quad-shaped token is 3.3.0.0, a product version from Declare the product version once, and pin every reader to it (#3222) #3225's own text, and the check asserts that the only such token is that known literal in version context rather than waiving the pattern -- a new address would still red.

One red attempt, diagnosed rather than re-run on hope

Darling whole-tree guards failed on attempt 1 of the final head with Darling.Tests Total: 8386, Errors: 0, Failed: 2, Skipped: 337, Not Run: 0 -- two DarlingDeployRollbackRetentionTests cases, both reporting "powershell.exe did not exit within 60 seconds running the extracted functions". Those tests extract functions from the deploy script and run them through powershell.exe; the diff contains CHANGELOG.md and nothing else, so there is no path from it to a PowerShell runner timeout.

Three facts placed it before the re-run rather than after: the same job was green on both earlier heads carrying 33 of these 35 lines; #3230's own head was 8/8 green including this job, so the merged content is not implicated either; and the only delta between the last green head and the red one is that merge plus two CHANGELOG.md lines. Attempt 2 passed in 231 s. Recorded rather than quietly re-run, because a red attempt in the history is something a later reader will otherwise have to re-diagnose.

Base

Cut from origin/dev at 5b6ad87a9 (#3231) in a fresh worktree. origin/dev then moved once, when #3230 merged, and it was merged in rather than rebased, so e45953960 stays intact in the history and #3230's entry lands in a commit of its own on top -- an entry omitted because it arrived mid-batch is how a backlog restarts, and the batch's purpose is to make [Unreleased] current at its own merged base.

Four commits, and each later one is a new commit rather than an amend: the earlier heads' eight checks had already run green, and rewriting them would have meant a force-push.

[claude]

Reviewed. This is a CHANGELOG.md-only PR (33 insertions / 1 deletion), so the T-SQL style, Lite/Darling parity, and security-review criteria don't have surface area to bite on here — no code changed.

Verified against the actual repo state:

• The file has exactly 1,246 link-reference definitions and no duplicates (grep -oE "^\[#[0-9]+\]" | sort | uniq -d is empty), matching the PR's claim.
• All 15 new definitions ([The store probe is 52.7% of query_store's wall time and 98.6% of what it examines is already stored: 31M references a day to find 1.4% missing #3189], [SerialLoopSeconds' chain bound is enforced against a watchdog that does not clock the serial loop #3204], [MCP reads return 16 server-local timestamps unmarked, beside naive-UTC fields in the same payload (#3198's siblings) #3206], [Desktop clock frames are wrong in opposite directions in both SKUs, and each SKU is right where the other is wrong #3207], [No guard scans a payload or render projection for a server-local column handed to a UTC-assuming consumer #3208], [ComposeCompiler has no duplicate-parameter-bind guard: a second bind of the same value shifts every later ordinal and 273 tests pass #3211], [Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is ever built, so it is reachable where Entra MFA is not #3214], [Six SQL-text pins assert a behaviour by grepping for its current spelling, so a semantically neutral rewrite reds four untouched files #3217], [Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is built (#3214) #3218], [Azure.Identity is unpinned and transitive, so #3218's broker-free guarantee rests on a dependency-graph fact nothing tracks #3219], [[BUG] Lite query and procedure grids render server-local DMV timestamps through the UTC formatter #3221], [Four shipped projects carry independent <Version> literals; the release gate reads a deprecated one and the nightly reads a different one #3222], [Weld concatenated literals before scanning store SQL, and make the PostgreSQL parse-check population what the reader files declare #3223], [Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226], [Scope the de-skew's accuracy claim where the column outlives a DST transition #3231]) are present exactly once.
• Spot-checked several entries against the code they describe (ServiceCommandDeadlines.SerialLoopSeconds's doc comment for the SerialLoopSeconds' chain bound is enforced against a watchdog that does not clock the serial loop #3204/Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199 correction, the Azure Existing Sign-In auth mode for Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is ever built, so it is reachable where Entra MFA is not #3214) — both match what's actually in the tree.
• Re-derived the merge order of the ten underlying PRs (De-skew the sixteen server-local timestamps MCP reads return, in both SKUs (#3206) #3212, Single-source the Query Store liveness touch guard and derive its width from the margin it spends (#3189) #3216, Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is built (#3214) #3218, Guard a server-local column reaching a UTC-assuming consumer, with the census (#3208) #3220, Weld concatenated literals before scanning store SQL, and make the PostgreSQL parse-check population what the reader files declare #3223, Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224, Declare the product version once, and pin every reader to it (#3222) #3225, Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226, Pair every desktop clock render with its column's frame, in both SKUs (#3207, #3221) #3227, Scope the de-skew's accuracy claim where the column outlives a DST transition #3231) via gh pr view --json mergedAt and confirmed both the Added and Fixed sections are in newest-merge-first order, consistent with the file's existing convention.
• The Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199 arithmetic-claim deletion is a clean edit (sentence boundary reads fine before/after), and the removed claim is independently corroborated as unsound by the doc comment already sitting on ServiceCommandDeadlines.SerialLoopSeconds in the current source.

One real issue found — left as an inline comment on line 13: the Added entry for the Azure credential-recording feature cites [#3218], but that number belongs to a different, already-separately-cited PR (the auth-mode change, cited via [#3214] in the entry above it). The actual implementing PR is #3224, which never gets a citation or definition anywhere in the file — contradicting this PR's own claim of full coverage over the ten merged PRs. Recommend citing [#3224] instead (no issue exists for it, so per the PR's own citation rule it should get its own PR-number citation).

[claude]

Reviewed. This PR only touches CHANGELOG.md (33 insertions / 1 deletion), so the T-SQL style / Lite-Darling parity guidance in CONTRIBUTING.md doesn't apply here — there's no code, no collectors, no C#. I focused the review on verifying the PR description's own claims against the actual GitHub state, since the description makes very specific, checkable assertions about the merge range, citations, and ordering.

Diff mechanics check out. Cross-checked against the GitHub API:

• The 10 PRs' mergedAt timestamps (De-skew the sixteen server-local timestamps MCP reads return, in both SKUs (#3206) #3212, Single-source the Query Store liveness touch guard and derive its width from the margin it spends (#3189) #3216, Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is built (#3214) #3218, Guard a server-local column reaching a UTC-assuming consumer, with the census (#3208) #3220, Weld concatenated literals before scanning store SQL, and make the PostgreSQL parse-check population what the reader files declare #3223, Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224, Declare the product version once, and pin every reader to it (#3222) #3225, Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226, Pair every desktop clock render with its column's frame, in both SKUs (#3207, #3221) #3227, Scope the de-skew's accuracy claim where the column outlives a DST transition #3231) match the newest-merge-first ordering used in both the ### Added and ### Fixed sections, entry by entry.
• All 15 new [#N] link definitions resolve, are cited by an added entry, and land at the correct ascending position ([The store probe is 52.7% of query_store's wall time and 98.6% of what it examines is already stored: 31M references a day to find 1.4% missing #3189] inserted mid-run between [The grid's two ceiling constants are maxima over open populations, so both go stale by construction — one moved three times in an evening, the other is now 3.6x too high #3188]/[query_store's sql_duration_ms includes store-side probe time, so the target-vs-store split points the wrong way on the most expensive collector #3192]; the other 14 appended at the foot).
• Desktop clock frames are wrong in opposite directions in both SKUs, and each SKU is right where the other is wrong #3207/No guard scans a payload or render projection for a server-local column handed to a UTC-assuming consumer #3208/Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is ever built, so it is reachable where Entra MFA is not #3214/Azure.Identity is unpinned and transitive, so #3218's broker-free guarantee rests on a dependency-graph fact nothing tracks #3219 correctly resolve as issues (not PRs) via the API, matching their /issues/N definitions; Weld concatenated literals before scanning store SQL, and make the PostgreSQL parse-check population what the reader files declare #3223/Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224/Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226/Scope the de-skew's accuracy claim where the column outlives a DST transition #3231 correctly resolve as PRs, matching their /pull/N definitions.
• Render every desktop timestamp through the renderer for its own clock frame (#3207) #3228 and Scope the de-skew's accuracy claim where the column outlives a DST transition #3229 (candidate duplicates of Pair every desktop clock render with its column's frame, in both SKUs (#3207, #3221) #3227 and Scope the de-skew's accuracy claim where the column outlives a DST transition #3231) are CLOSED, not merged — correctly excluded. Say that the recorded-size read's cheapness is incidental, and on which axis that differs (#3199 follow-up) #3215 merged before Record the CHANGELOG entries for thirty-three changes across twenty-six merged pull requests #3213's own cutoff — correctly excluded as already covered by the prior batch.
• The single deletion is exactly the arithmetic clause the PR body describes removing from the Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199 entry; the surviving sentence is otherwise untouched.

One real problem: the "#3230 is still open" claim is now false, and this batch is missing its entry.

The PR body states twice that #3230 was excluded because it's still open ("#3230 is still open and is not absorbed" / "#3230 was open at the cut and still open at push"). Per the GitHub API, #3230 ("Report the member ranges that stop short of their own content") merged to dev at 2026-09-09T19:31:55Z. This PR's own commit timeline is:

• commit 1: 2026-09-09T19:26:44Z
• Report the member ranges that stop short of their own content #3230 merges: 19:31:55Z
• commit 2: 2026-09-09T19:36:27Z

So #3230 had already been on dev for ~4.5 minutes by the time the second commit was pushed — the "still open at push" premise doesn't hold for the push that actually shipped. Since this PR's own stated boundary is "ten pull requests merged to dev after #3213," and #3230 merged after #3213 (and even after #3231, the PR's own stated cut point), it's in-range and has no entry here.

Given the file's own rule that "lanes do not edit CHANGELOG.md" and batches are periodic, this could just roll into the next batch — but as written, the PR's stated verification ("nothing merged inside the range is uncovered") is factually incorrect as of the second commit, and whoever runs the next batch needs to know to pick up #3230 specifically (it merged in the gap between this PR's two commits, which is an easy case to miss if the next batch's range-detection just looks at "after #3232's base"). Worth either adding #3230's entry now or flagging it explicitly for the next pass rather than leaving the stale "still open" claim as the record.

🤖 Generated with Claude Code

[claude]

Reviewed. This PR only touches CHANGELOG.md (35 insertions / 1 deletion), so the T-SQL style and Lite/Darling parity checks don't apply to the diff itself — no source files changed.

Independent verification performed (without relying on the PR description's own claims):

• No duplicate link-reference definitions: grep -oE '^\[#[0-9]+\]:' CHANGELOG.md | sort | uniq -d → empty.
• Citation/definition balance: distinct ^[#N]: definitions and distinct [#N] citations both count 1,247 in the resulting file — consistent with "no dangling citations, no number defined twice."
• Line count: file is exactly 3,591 lines post-patch, matching the PR body's claimed blob line count.
• Citation-to-source mapping spot-checked against the actual GitHub API (not just trusted from the description) for all 11 PRs in the batch (De-skew the sixteen server-local timestamps MCP reads return, in both SKUs (#3206) #3212, Single-source the Query Store liveness touch guard and derive its width from the margin it spends (#3189) #3216, Add a broker-free Entra sign-in path: ActiveDirectoryDefault returns before the WAM broker is built (#3214) #3218, Guard a server-local column reaching a UTC-assuming consumer, with the census (#3208) #3220, Weld concatenated literals before scanning store SQL, and make the PostgreSQL parse-check population what the reader files declare #3223, Record which credential DefaultAzureCredential selected, on an event-id allowlist (#3218 follow-up) #3224, Declare the product version once, and pin every reader to it (#3222) #3225, Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226, Pair every desktop clock render with its column's frame, in both SKUs (#3207, #3221) #3227, Report the member ranges that stop short of their own content #3230, Scope the de-skew's accuracy claim where the column outlives a DST transition #3231) — every issue-vs-PR-number citation choice matches what each PR's own title/body says it closes, including the multi-entry PRs (Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226 → 5 entries citing ComposeCompiler has no duplicate-parameter-bind guard: a second bind of the same value shifts every later ordinal and 273 tests pass #3211/SerialLoopSeconds' chain bound is enforced against a watchdog that does not clock the serial loop #3204/Six SQL-text pins assert a behaviour by grepping for its current spelling, so a semantically neutral rewrite reds four untouched files #3217/itself/Azure.Identity is unpinned and transitive, so #3218's broker-free guarantee rests on a dependency-graph fact nothing tracks #3219; Pair every desktop clock render with its column's frame, in both SKUs (#3207, #3221) #3227 → 3 entries citing Desktop clock frames are wrong in opposite directions in both SKUs, and each SKU is right where the other is wrong #3207/[BUG] Lite query and procedure grids render server-local DMV timestamps through the UTC formatter #3221).
• Ordering: both ### Added and ### Fixed new entries are strictly newest-mergedAt-first, verified against each PR's real merge timestamp via gh pr view.
• The one edit (striking the 10 sites × N < 60 s arithmetic claim from the Store disk-pressure check's pg_database_size times out in bursts: 5s deadline sized on a 4.05 GB store, beating against the 300s checkpoint cycle #3199 entry) matches the before/after quoted in the description exactly, and is corroborated by Five instruments that assert something other than what they check (#3211, #3204, #3217) #3226's own description explaining why both operands were wrong.
• PR is currently MERGEABLE / UNSTABLE — a couple of checks (review, verify, Darling whole-tree guards) show no conclusion yet at review time; worth confirming they finish green before merge, though nothing here should touch code-level guards since no source file is in the diff.

No correctness, security, or parity issues found in the diff.

[claude]

Reviewed. This PR touches only CHANGELOG.md (35 insertions, 1 deletion) — no T-SQL, no Lite/Darling C# code, no config. The style/parity/security review categories in the brief don't apply here since no code changed.

Spot-checked the claims in the PR description against the actual diff and file state:

• Citation/definition integrity: every new bracket citation added ([#3230], [#3207], [#3219], [#3224], [#3208], [#3214], [#3231], [#3221], [#3211], [#3204], [#3217], [#3226], [#3206], [#3222], [#3223], [#3189]) has a matching link-reference definition, and vice versa. No duplicate [#N]: definitions.
• #3218 handling: confirmed there is no [#3218] bracket anywhere in the file (citation or definition) — it was correctly demoted to plain prose (a follow-up to #3218) in the #3224 entry rather than left as an orphaned link.
• #3223 and #3219 double-leads: confirmed #3223 leads both of its entries (the two guard fixes) and #3219 is referenced 3 times (1 definition + 2 citations, from #3226's dependabot note and #3224's entry), matching the description.
• #3199 correction: confirmed the old sentence containing the 10 sites × N < 60 s arithmetic claim is removed and the surrounding text is otherwise byte-identical; [#3199]'s definition itself is untouched (pre-existing, still points at /issues/3199).
• Definition ordering: [#3189] is correctly inserted ascending between [#3188] and [#3192]; the new run at the file's end (3204...3231) is ascending.
• PR vs. issue citation choice: verified via gh pr view that #3223, #3224, #3226, #3230, #3231 are actually pull requests (not issues), consistent with using /pull/N link targets for those five and /issues/N for the rest.
• Counts match: 6 new Added entries, 12 new Fixed entries, 16 new definitions, 35/1 insertions/deletions — all as claimed.

No factual or formatting problems found. Nothing to flag for Lite/Darling parity, T-SQL conventions, or security since no code is in this diff.