191 merged PRs carry a CHANGELOG entry in the PR body that never reached the file

[gsdali]

python3 Scripts/check-changelog-transcription.py --verify-transcribed, run on main at
24364cef with #1680 applied:

Merges audited since 03bf315a: 367
  entry transcribed:      46
  explicitly opted out:   2
  NO entry, no opt-out:   319

Verified against each PR body (#788):
  transcribed late, nothing to do:  68
  entry in the PR body, NOT in the CHANGELOG:  191
  PR body has no entry section at all:  32
  PR body says None, by design:  26
  could not verify:  1

#1680 cleared the most recent 42 (#1590 through #1676). 191 remain, spread from #1118 to #1560:

PR range	Count
#1100s	23
#1200s	32
#1300s	60
#1400s	42
#1500s	33
#1600s	1

Why this blocks v4.0.0

Step 2 of the release process in CLAUDE.md is "transcribe the CHANGELOG; at release you check the
## Unreleased section is complete". It is not complete: 191 shipped changes have no release note,
including the bulk of the #1413 bridge-correctness sweep (the #1300s and #1400s) and most of the
#1551 Swift-layer sweep (the #1500s). Several are MAJOR, and
semver-at-release assembles docs/SEMVER.md from the same PR
bodies, so it inherits the same gap.

What the work is

Not a reconstruction. The text is in each PR body forever, which is the property
changelog-on-merge relies on. It is a lookup, a
repunctuation pass for em-dashes, and placement in merge order. #1680 is the worked example for 42
of them, tooling included: fetch the bodies, extract the ## CHANGELOG entry block, insert at the
right merge-order anchor, then re-run --verify-transcribed and confirm "transcribed late" rose by
exactly the number transcribed.

Two things #1680 learned that the next batch should carry:

• The audit over-reports. docs(#1399): read the booleans family of the unlaned refman-coverage lane #1649's entry was in the file; the audit called it missing because the
  PR body's heading writes #1399 as a markdown link and the file's writes it bare. Match on the
  normalised heading (links flattened, backticks stripped) before believing any single row.
• A batch this size wants one PR per few dozen entries, not one PR for 191, and the batch route
  requires no other open PR touching docs/CHANGELOG.md, so they go in sequence.

The separate 32

PR body has no entry section at all is a different job: there is nothing to transcribe and the
entry has to be written from the merged diff. #1680 did seven of those and named each one in its
body. Worth splitting out once the 191 are done.

Related

• The CHANGELOG-on-merge policy has no automated backstop; a forgotten transcription is silent #742, the audit script itself.
• Nine merged PRs' CHANGELOG entries were never transcribed, and repunctuating a banned em-dash defeats the audit #1125 / PR docs(#1125): transcribe nine missed CHANGELOG entries, and stop the audit reporting a correct one #1126, the last catch-up pass, which fixed the substring matching that made a
  repunctuated heading read as never transcribed.

[gsdali]

Done, in four PRs: #1684 (#1118-#1235, 47), #1685 (#1236-#1357, 48), #1686 (#1358-#1465, 47),
#1687 (#1466-#1562, 48). 190 entries transcribed, the 191st being #1649, which was never
actually missing.

Result

check-changelog-transcription.py --verify-transcribed on main at 1a88421e:

	when this issue was filed	now
transcribed late, nothing to do	68	258
entry in the PR body, NOT in the CHANGELOG	191	1
PR body has no entry section at all	32	32
PR body says None, by design	27	27

## Unreleased goes from 122 entries to 312. The file is 561 ### headings, up exactly 190,
and still at zero em-dashes.

Method, since the next pass will want it

Placement is by merge order, not PR number. ## Unreleased is newest-first, so each entry goes
immediately above the first existing entry that merged before it. The merge index comes from
git log --reverse over origin/main; existing headings are matched back to the PR that wrote them
by normalising the heading (flatten markdown links, strip backticks, collapse whitespace) and
looking it up across all 554 PR bodies in the audit window. 104 of the 122 existing headings
resolved that way, which was enough: an unresolved heading is simply skipped as an anchor.

Every insertion is pure. All four diffs are insert-only, 1734 added lines and zero removed
in total. No existing entry moves relative to another and none is reformatted. Verified by holding
three invariants constant across all four PRs: triple-blank-line runs (21), the count of ### with
no blank line before it (1, the pre-existing v1.11.3 release-history heading), and the set of
adjacent out-of-merge-order pairs (18, all pre-existing). The 190 insertions introduced zero new
ordering violations.

Verbatim is checked, not assumed. After repunctuation, every one of the 190 entries is present
in the file as an exact substring of its PR body. Checked programmatically per batch, not by eye.

The audit delta is the real check. Each batch was required to move "transcribed late" up and
"NOT in the CHANGELOG" down by exactly its own size: +47/-47, +48/-48, +47/-47, +48/-48. A
placement bug that dropped or duplicated an entry would show there.

Repunctuation

30 em-dashes across 23 entries, each replaced individually per
okf/policies/writing-style.md: 7 in batch 1, none in batch 2, 7 in batch 3, 16 in batch 4. Every
replacement was asserted unique within its own entry before being applied, and each is confirmed to
appear exactly once in the file afterwards.

One needed care. #1467's em-dash is in the heading, and the audit's own #1125 fallback matches
the part of a heading before its first em-dash as a substring, so a repunctuation that ate text on
either side of the dash would make a correctly-transcribed entry read as missing forever. Checked
explicitly: no-op — now actually perform became no-op, and now actually perform, the fallback
prefix is 70 distinctive characters, and it is still a prefix of the transcribed heading.

Confirmations of the two warnings in this issue

• The audit over-reports, exactly once. docs(#1399): read the booleans family of the unlaned refman-coverage lane #1649 was the only false positive in 191. Normalising
  headings before believing a row caught it, as predicted.
• Six entries are a heading and nothing else (fix: merge duplicate Curve-Surface Intersection test suites (#1291) #1338, fix: merge duplicate Surface-Surface Intersection test suites (#1292) #1345, fix: remove duplicate curvatureOfCircle test, tighten circleRadius tolerance (#1293) #1347, fix: remove reference to non-existent Issue595DeprecatedLocalCurvatureTests (#1294) #1348, fix: factor void-shape and zero-size-vertex fixtures into shared helpers (#1295) #1349, fix: factor five-of-six-faces open-shell fixture into shared helper (#1296) #1350). Their
  PR bodies really do hold only the ### line. Transcribed as written, and called out in docs(#1681): transcribe the CHANGELOG entries for #1236 through #1357 #1685's
  body so they read as thin rather than truncated in transit.

Follow-up

The separate 32 are filed as #1688. Ten of them already have an entry that somebody drafted
later, so the real remainder is 22, and roughly a third of those look like they owe a
No-Changelog: trailer rather than an entry.