Record the CHANGELOG entries for sixty-five merged changes

[erikdarlingdata]

Appends the collected [Unreleased] entries for the 85 pull requests merged to dev from 8210012bc (the previous batch, #2984) through #3147. 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 PR's own description instead. This is the periodic commit that applies it.

What is in the diff

CHANGELOG.md only, 141 insertions and no deletions: 70 entries prepended inside ## [Unreleased] — 10 under ### Added, 13 under ### Changed, 47 under ### Fixed — plus the 71 link-reference definitions those entries need, appended to the block at the foot of the file. Entries are ordered newest-merge-first within each section, matching the file's existing order. Nothing already in the block was edited, reworded or reordered.

Base

origin/dev moved twice while this was assembled. Both were merged in rather than rebased, so 33211034e is intact in the history: #3149 landed during the survey and #3147 landed after it closed, arriving as the merged base. Both carry entry text and both are included, #3147's 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. There is no PR in the range whose entry this diff knowingly lacks.

Where the text came from

85 pull requests in the range. 66 contributed an entry, 19 did not.

Of the 66, 54 carried their entry text under a heading in their own body, one (#3033) carried it under a bare paragraph rather than a heading, and 11 carried it in the lane's handoff file rather than the PR body (#3040, #3042, #3046, #3075, #3077, #3078, #3079, #3080, #3084, #3085, #3086). The text is used close to verbatim; where a lane wrote a plain bullet, its own opening clause is bolded to match the shape every existing entry in the block uses, and a redundant Fixed: / Added: prefix is dropped where the section heading already says it. Three PRs contributed more than one entry because they fixed more than one thing: #3027 (two), #3056 (three, across ### Fixed and ### Added), #3117 (two).

Two entries needed a choice recorded here. #3077's lane offered a first draft and a post-review final; the final is used, because the draft cites a count of inadmissible readings the lane says a third reading falsified mid-review. #3079 and #3085 each offered a full entry and a one-liner; the full form is used.

Section placement follows the block's existing treatment of guard work rather than a rule invented here: a guard that never existed is Added (#3081, #3083, #3049's grid check), a guard whose reach was widened as policy is Changed (#3063, #3067, #3074, #3076, #3082 — the shape [#2490]'s entries already sit under Changed), and a guard that gave a wrong answer is Fixed (#3052, #3088, #3094, #3140 — the shape [#2927]'s entry already sits under Fixed).

The 19 with no entry

Two say so themselves: #3043 ("a comment aligning with a naming convention is not a user-visible change and does not earn a line") and #3020 ("no behaviour change — every edit is inside a comment or a changelog line"; it is also the only PR in the range that touched CHANGELOG.md, a one-line placeholder scrub already on dev).

Seventeen have no entry text in the PR body, in the lane handoff files, or anywhere else searched, and are not covered by the existing block: #2982, #2983, #2985, #2986, #2987, #2988, #2989, #2990, #2992, #2998, #3000, #3001, #3002, #3003, #3005, #3007, #3025. All 17 merged between 22:44 on 09-04 and 10:34 on 09-05, before the entry-rides-the-description routing settled. Several are user-visible (#2985 takes the default trace's archival-empty cutoff from the server's clock, #3000 screens rdsadmin out of per-database enumeration, #3003 bounds pg_index_bloat by bytes, #3007 derives the viewer's interactive read deadline from its fan-out width). No entry is invented for them here — these entries cite measurements only the lane that did the work can vouch for — so they want a backfill in the shape of #2260.

How the count was checked

Two ways that fail differently. git diff reports 70 added - lines and 71 added link definitions, totalling the 141 in the diffstat; independently, re-parsing the written file and subtracting the per-section counts measured before any edit (96 / 15 / 143) gives a delta of exactly 10 / 13 / 47. Separately, every one of the 66 contributing PRs is asserted to have its issue reference present among the added lines, and every added line is asserted to be non-trivial and house-shaped rather than merely present.

The extractor was also run twice. A first pass that treated any heading as a section terminator lost #3149's entry completely, because that entry sits inside a fenced block whose first line is ### Fixed; a fence-aware pass recovered it and agreed with the first pass on all 53 others; #3147's entry was extracted with that same fence-aware pass. [Cc]hangelog does not match CHANGELOG, and requiring CHANGELOG at a heading's start misses #3122 ("Proposed CHANGELOG entry"), #3146 ("Changelog entry for the coordinator") and #3117's ###-level heading, so the search is case-insensitive, any level, and matches mid-line.

CHANGELOG.md is outside FleetIdentifierScrubTests' population by construction, so CI does not scan this text for identifiers. The diff was scanned by hand against the guard's own two patterns and for RDS endpoints, ARNs, 12-digit account ids, dotted-quad and IPv6 addresses, and the fleet's own naming: clean. The one address-shaped token, 2001:db8::1, is the RFC 3849 documentation prefix in #3030's account of a log-prefix regex; 01:00:00 and 24:00:00 are PostgreSQL interval literals.

[claude]

Reviewed. This PR touches only CHANGELOG.md (139 insertions, no deletions) — no application, T-SQL, or config code, so the T-SQL style, Lite/Darling parity, and security-boundary checks in the review brief don't have a surface to apply to here. Verified the mechanics of the batch instead:

• All 70 newly added [#NNNN]: link-reference definitions are used by at least one entry, and every [#NNNN] reference used in the 69 new bullets (including the pre-existing #1586/#1776/#1778/#1902/#1937 refs) resolves to a definition somewhere in the file. No duplicate link definitions anywhere in CHANGELOG.md.
• Entry/link counts match the PR description's own accounting: 69 - ** bullets (10 Added / 13 Changed / 46 Fixed) and 70 appended link definitions.
• Spot-checked several PR→issue mappings the description calls out as sourced from lane handoff files (Fixes #3069 #3077→HeaviestHourlyRefreshObservedCeilingSeconds is one sample of a twelve-hour-old regime, and #3035's grid plus a build-time assertion are sized against it #3069, Fixes #3076 #3079→CrossAppGuardCiGateTests cannot see a cross-app read spelled as a bare directory name, and there is a live one #3076, Fixes #3074 #3080→Replace the cross-app guard's hand-rolled MSBuild parsing with dotnet msbuild -getItem, which is offline, 0.27s, and fully evaluated #3074, Fixes #3082 #3084→Fourth cross-app spelling: the C# stage needs a root literal as Path.Combine's first argument, so a specific-file read with a variable base is invisible — and four in one night is the finding #3082, Fixes #3081 #3085→CONTRIBUTING.md's T-SQL conventions are unenforced: a COUNT() violation passed every gate, caught only by a review bot #3081, Fixes #3013 #3040→Under store contention the alert pass fails first and silently, so the self-alert that would report it is the first thing lost #3013, Collect the store's own PostgreSQL log as a self-monitoring census (#3021) #3046→The store's own PostgreSQL log is invisible to the product — collect it as a self-monitoring source #3021, Fixes #3067 #3075→CrossAppGuardCiGateTests anchors on Lite/, so the two Darling.Tests reads of Lite.Tests/ are invisible to it #3067) against the actual PR titles via gh pr view — each resolves to the issue number the corresponding changelog entry links to, so the apparent "missing" PRs are accounted for correctly (entries link the fixed issue, not the fixing PR).
• New definitions are appended in ascending numeric order, consistent with the existing convention at the foot of the file, and the ### Added / ### Changed / ### Fixed structure under [Unreleased] is unchanged in shape.

No issues found.

[claude]

Reviewed. This PR touches only CHANGELOG.md (139 insertions, no code changes), so the T-SQL style, Lite/Darling parity, security, and performance criteria don't apply here — there's no source to check them against.

Verified the mechanical part of the aggregation:

• Every [#NNNN] reference used in the file has exactly one matching [#NNNN]: url link definition, and vice versa (1203 used == 1203 defined, bijective) — no dangling references, no orphaned definitions.
• No duplicate link-reference definitions were introduced (checked the full file, not just the diff).
• All 70 new definitions line up correctly with the entries that cite them.

No issues found.

[claude]

Reviewed against the T-SQL style, Lite/Darling parity, security, correctness, and performance criteria — none apply here, since the diff is CHANGELOG.md only (141 insertions, 0 deletions): no application code, no T-SQL, no Lite or Darling source touched.

Structural checks performed on the resulting file:

• All 1,204 unique [#NNNN] reference tokens in the file resolve to a matching [#NNNN]: https://... definition line, and there are no duplicate definition lines.
• No unintentional duplicate bullet entries — the only PR numbers appearing on more than one bullet (Fixes #3017 (items 1 and 2) #3027, Fixes #3049 #3056, Fixes #3111 #3117, plus the two The migration advisory-lock wait cannot cover an arbitrary number of data-moving rungs #2894/Two COPY fault-reporting follow-ups from #3110: three sibling sites keep the unclosed start-phase residual, and the PostgreSQL-target timeout sentence can be applied to a store write #3111 follow-ups) match what the PR description already calls out as multi-entry PRs.
• Section placement (Added/Changed/Fixed) and ordering (newest-merge-first) is consistent with the rest of the block.

No findings. This is a mechanical aggregation of already-reviewed PR entries, so I did not re-litigate the substance of those 66 individual changelog blurbs — the review scope for this PR is the aggregation itself, which looks correct.

[erikdarlingdata]

Claude posting for Erik Darling

Correction to the merged commit's own numerals, recorded here because the commit is immutable and this is the nearest durable surface to it.

The squash subject on dev (11a5e53ee) reads "sixty-five merged changes", and its body's first bullet carries the figures as they stood before #3147's entry was added: 69 entries, 10 Added / 13 Changed / 46 Fixed, 84 pull requests, 70 definitions.

The artefact is right; only those numerals are stale. What actually landed:

entries	70 (10 Added / 13 Changed / 47 Fixed)
contributing PRs	66
range surveyed	85 PRs, 8210012bc..5d5b77cc8
link-reference definitions	71
diff	141 insertions, 0 deletions, CHANGELOG.md only
partition	66 contributed + 2 no-entry-by-own-statement + 17 needs-a-human = 85, and git log over that range yields exactly 85 PR-numbered commits with 0 unnumbered

The description above is correct on all of these; the second commit message in the squash body also records #3147's entry being added "with the one link-reference definition it needs", so 69+1 and 70+1 are reconstructable from the message itself.

Deliberately not editing the title. Retitling to "sixty-six" would make this page disagree with the commit it produced, and a title/commit divergence is worse for anyone tracing 11a5e53ee back than a stale-but-consistent pair with an accurate correction adjacent to it. No CHANGELOG change is warranted either — this is a commit-subject numeral, not a claim inside the file.

How it happened, since the mechanism generalises. When #3147's entry was authorised the description was re-audited numeral by numeral against the artefact — 20 claims, all passing. But that audit's population was the PR body and nothing else. The PR title and the commit messages were never in it, and GitHub derives the squash subject from the title. So a stale figure sat two surfaces away from a thorough check that structurally could not fail on it.

That is the same defect class this batch's own entries keep describing: a check shaped like the answer, passing inside its population while identical figures go unverified outside it. The earlier catch in this PR — a description claiming "four PRs contributed more than one entry" while listing three — was luck of population rather than rigour, because that numeral happened to live in the body.