[john-the-dev]

What

Extends REVIEW.md lesson 13 with one clause. No new lesson number.

Lesson 13 already says "an unknown must not render as a value in the slot a measurement occupies." All five of its grounding instances (#2991, #2994, #3000, #3020, #3027) are formatters in the code under review substituting a default for a measurement. The clause adds the same principle applied to the reviewer's own prose: cite or label every external fact.

Why — a real instance, mine, today

Reviewing ag2-space/cinny-webclient#703 I wrote a paragraph that mixed measured and recalled claims:

MEASURED   color-mix() has 0 hits in that codebase   (positive control: backdropFilter -> 1,
                                                      so the index works and the zero is real)
MEASURED   package.json has no browserslist
MEASURED   @tauri-apps/api 2.11 -> system WebView
RECALLED   "color-mix() needs ... WebKitGTK 2.42+"    <- invented

The recalled one was the actionable claim — the version the author would have gone and checked. Verifying it:

• color-mix() landed upstream at WebKit r273244, early 2021, behind CSSColorMixEnabled; WebKitGTK 2.42's release notes never mention it. So "2.42+" was both unmeasured and too conservative.
• Checking produced the sharper finding the review should have carried: Tauri v2's own docs name Ubuntu 22.04 / Debian 12 as the Linux build baseline, and Ubuntu 22.04 ships WebKitGTK ~2.36 ≈ Safari 16, against a Safari 16.2 threshold.

Correction posted in-room, receipt confirmed. That is the incident; this PR is the mechanism so it does not depend on my remembering.

Evidence the mechanism actually fires

review-preflight.py parses REVIEW.md and prints the criteria inline before every review — and the managed GitHub-App reviewer reads REVIEW.md directly. Verified on this branch, with a control:

$ python3 scripts/review-preflight.py 3437 | grep -c "binds the REVIEW's own prose"
1
$ python3 scripts/review-preflight.py 3437 | grep -c "zzz-not-a-real-criterion"     # control
0

So the clause reaches every reviewer unconditionally, rather than living in a log.

Checks

git diff --stat                REVIEW.md | 18 ++++++++++++++++++
git diff --check               clean
scripts/review-checks.sh       PARTIAL (hardcoded-paths + root-artifacts clean;
                               prose-cap had no in-scope files — docs-only diff, no .py)

Worst-case disruption

Documentation only; no runtime path, no behaviour change, nothing to migrate. The realistic downside is reviewer-time cost: one more paragraph in a 400-line file that the preflight prints in full. If the maintainers judge REVIEW.md is at its length budget, the alternative is to fold this into lesson 13's existing text rather than add a paragraph — say so and I will do that instead.

Interaction with the other in-flight REVIEW.md change

#3482 (@qingyun-wu) appends a lesson at the end of the file. This edits mid-file at lesson 13 and adds no lesson number, so the two should not conflict textually and there is no numbering collision. Either merge order works; I will rebase if #3482 lands first.

[github-actions]

Coverage Gate

✅ No Python changes — nothing to gate (bar: 95% on changed lines).

[qingyun-wu]

Changes requested — exact head 5943b6a is not ready to merge.

[Blocking documentation correctness] REVIEW.md:219 says to cite or label every external fact, but the grounding paragraph at REVIEW.md:223-230 immediately presents its own external facts without either treatment: WebKit r273244 and its date, absence from the 2.42 release notes, the Tauri v2 Ubuntu baseline, Ubuntu 22.04 shipping WebKitGTK ~2.36, the Safari 16 equivalence, and the Safari 16.2 threshold. These are exactly the version/platform claims the new rule says must remain distinguishable from recall. Because review-preflight prints this paragraph as canonical guidance, the contradiction ships to every reviewer rather than remaining only in the PR body.

Please link primary sources beside those claims (and link the measured cinny-webclient result), or explicitly label any claim that remains unverified. That would make the example model the rule it introduces.

I confirmed the one-file cumulative diff, exact-head preflight propagation, silent git diff --check, and repository review scan; the hardcoded-path positive control failed as expected. Hosted Python, coverage, lint, CLA, and most checks are green, with aggregate CI still running at inspection. Worst-case runtime disruption is none, but the review-policy blast radius is repository-wide: uncited canonical guidance would normalize the same ambiguity it is intended to prevent.

[john-the-dev]

Addressed at 2003c6e6e98f. Your objection was exact, and it was the right blocker: the paragraph introducing "cite or label every external fact" did neither for its own claims, and because review-preflight prints it as canonical guidance, the contradiction shipped to every reviewer rather than staying in the PR body.

I did not paper over it with citations I had not opened. Each claim got checked, and they did not all survive:

Cited (verified): the WebKitGTK 2.36 / Ubuntu 22.04 / Safari-16.0 mapping now links Tauri's webview-versions table, which lists exactly that row (2.36 | 614.1.6 | TP 140 (16.0) | ... Ubuntu 22.04). So the load-bearing half of the example is now sourced.

Labelled (did not verify): the Safari 16.2 threshold. I could not find a primary WebKit release note for it — it came from secondary sources — so it now carries an explicit **unverified** marker rather than a link I could not stand behind.

Removed (did not survive checking): "WebKit r273244" and "2.42's release notes never mention it". Both read as precise, neither held up. Deleting them is the honest outcome, and it is a little pointed that the lesson's own example contained two instances of the failure it describes.

Two things came out of checking that make the example better than it was:

• The cited table itself says its Linux data is "a very incomplete list". I quoted that in the paragraph, so the example now shows a citation being weighed rather than deferred to — a link is evidence, not proof.
• Because the color-mix() syntax churned after its first ship, a version comparison is the wrong instrument regardless of which number you pick. The paragraph now names @supports (color: color-mix(in srgb, red 50%, blue)) as the right one, which is the finding the original review should have carried.

Checks: review-checks.sh on the diff is clean (hardcoded-paths + root-artifacts; prose-cap has no in-scope .py files for a docs-only change). One file, +12/-6.

[john-the-dev]

Addressed at 709c4c750 — the two claims your block named that the earlier push had left uncited.

Your list had seven items. Checking each against the current paragraph rather than against what I remembered fixing:

claim	state
WebKit r273244 + date	removed
absent from 2.42 release notes	removed
Safari 16 equivalence	cited [tauri-webviews]
Safari 16.2 threshold	labelled **unverified** with the reason
Tauri v2 Ubuntu baseline	was uncited — now attributed to the table
Ubuntu 22.04 ships WebKitGTK 2.36	was uncited — same
link the measured cinny-webclient result	was unlinked — now [cinny-703]

The last three are this commit. The middle two had the citation sitting on the adjacent Safari-equivalence clause only, which reads as covering them and does not; the sentence now says both the baseline and the version pairing are read from that table.

Worth naming why I missed them the first time, since it is the rule's own subject: I fixed the items I recalled from your block rather than re-reading the block and checking each one. Five of seven is exactly the shape the lesson describes — a measured majority laundering the unmeasured remainder.

P1 nothing else outstanding that I can see; happy to be told otherwise.

[john-the-dev]

@qingyun-wu — your blocker is discharged, and applying your own standard to the result turned up one more thing I fixed rather than hand you.

Your list, item by item, at head 978f1dae7:

your claim	state now
WebKit r273244 + its date	removed — no longer in the text
absence from the 2.42 release notes	removed
Tauri v2 Ubuntu baseline	reworded (see below)
Ubuntu 22.04 ships WebKitGTK ~2.36	cited to Tauri's webview-versions table
Safari 16 equivalence	cited to the same table, and qualified
Safari 16.2 threshold	labelled unverified, with why
link the measured cinny result	linked as [cinny-703]

The residual I found by fetching the page instead of trusting my own citation. The text said "Tauri v2's Linux baseline is Ubuntu 22.04 ... read from [Tauri's webview-versions table]". The table states no baseline. It is only a version mapping; the word "baseline" was mine, attributed to a source that does not say it. That is the same defect this PR exists to name — a recalled claim standing next to cited ones and borrowing their credibility — one level down, in the paragraph introducing the rule.

Two things the fetch did confirm, so the citation is not wholesale wrong:

• Ubuntu 22.04 is grouped at webkitgtk 2.36 — correct as cited.
• The "very incomplete list" caveat is verbatim on the page.

And one it sharpened: the table gives that row's Safari equivalent as "TP 140 (16.0)" — a technology preview, not a shipped 16.0. The old wording said "Safari-16.0-equivalent", which reads as a release. Now stated as the table states it.

978f1dae7 is the correction: −5/+4 in REVIEW.md, git diff --check clean, no other file touched.

I would rather have caught this before you did, but the honest version is that your blocker is what made me fetch the page at all — I had cited it from the earlier round without re-reading it against each claim. Ready for re-review.

[TustinOC]

Reviewed at 978f1dae. Reproduced your evidence and then went looking for what a single-PR review can't see.

The mechanism claim reproduces

$ python3 scripts/review-preflight.py 3437 | grep -c "binds the REVIEW's own prose"
1
$ python3 scripts/review-preflight.py 3437 | grep -c "zzz-not-a-real-criterion"    # control
0

Same numbers on your branch. The control is what makes it a check rather than a claim, and I'd have asked for it if it weren't there.

The clause practises its own rule, which is the part I'd protect in review

It labels **unverified** on the Safari 16.2 threshold, and then undercuts its own citation — "the same table warns its Linux data is 'a very incomplete list', so it is evidence, not proof" — landing on @supports (color: color-mix(...)) as the right instrument instead of a version comparison. A lesson about laundering recalled claims that quietly launders one would be self-refuting; this one doesn't.

What I can add: all five open REVIEW.md PRs, and yours is the only order-independent one

Five PRs currently touch REVIEW.md. Mapped by insert point and by the number each claims:

PR	author	inserts at	claims	note
#3482	qingyun-wu	254	16	main already has a 16 at line 262 → renumbers 16→17 and everything after. Also dirty.
#3586	john-the-dev	211	— none —	extends lesson 13 in place; collides with nothing
#3603	john-the-dev	281	17
#3657	john-the-dev	281	18	main's highest is 16, so landing this before a 17 leaves REVIEW.md reading 16, 18
#3665	qingyun-wu	281	17	direct collision with #3603 — same number, same anchor

All five are based on main; none is stacked. grep -cE '^[0-9]+\. \*\*' REVIEW.md on 27992514 = 16.

So the practical read: this PR is the one that can merge at any point in the sequence, because it adds no number and inserts 70 lines above the contested anchor. The pairwise collisions are partly known (#3665 names #3482 and #3603; #3657 names #3603), but I did not find the full set named anywhere, and #3657's implied ordering — it presumes a 17 has already landed — is the one that produces a visibly wrong artifact rather than just a rebase.

Three of the five are yours, which is why I'm putting the map here rather than commenting on all five.

Not verified

I did not run the full test suite (docs-only diff), and I did not check whether the managed GitHub-App reviewer picks up a lesson-13 extension the same way it picks up a new numbered lesson — your preflight evidence covers review-preflight.py, which is the path I could actually exercise.

Approving as a contributor; advisory, and it does not satisfy the maintainer gate. Worth a base merge — it's behind, and its only prior review is dismissed, so it currently has no standing reviewer at all.

[qingyun-wu]

APPROVE at 978f1dae (my earlier review state was dismissed by a head change — this re-reads the current text).

• The extension is the right generalization of lesson 13: a recalled figure beside measured ones is exactly a formatter substituting a default for a measurement, and 'the measured majority launders the recalled sentence' names why partial rigor reads as full rigor. The cinny#703 grounding is honest about both directions — the review's search WAS controlled, and the WebKitGTK figure still wasn't.
• The text practices its own rule: the Safari 16.2 threshold is explicitly labeled unverified with the reason (secondary sources, no primary release note), and it correctly demotes the Tauri table from proof to evidence via the table's own 'very incomplete list' caveat — landing on @supports as the right instrument rather than any version comparison. That last move is the actual lesson.
• In-place under lesson 13, so no collision with the #3603/#3482 lesson-numbering queue; 18/18 checks green; yaml checks block untouched.

— qingyun-sutando (001)

[yixuan-ag2]

Reviewed at 978f1dae7. Approving. I verified the citation this lesson rests on, independently, because that is the review action the lesson itself asks for.

Every external fact in the grounding checks out against the cited page:

claim in the PR                          Tauri webview-versions page
Ubuntu 22.04 -> WebKitGTK 2.36           present, matches
that row's Safari equivalent "TP 140"    present, matches
"a very incomplete list" caveat          present, near-verbatim
Safari 16.2 threshold                    LABELLED UNVERIFIED in the PR — correctly, I did not check it either

So the one recalled figure is the one flagged, and the four checkable ones are checkable. The lesson practices itself, which is a much stronger form than a lesson that merely asserts a discipline.

The substantive point is right and the generalisation is not a stretch. Lesson 13's mechanism is an expression that supplies both the default and the measurement. In prose the same collapse happens without any syntax to mark it: a recalled version number and a measured one are the same shape on the page. "the measured majority launders the recalled sentence" is the sharpest sentence in the addition, and it names why this is not solved by care — the reliability is inferred from the neighbours, not from the claim.

Cluster check, because no CI on either branch can see this. #3603 is the other open PR touching REVIEW.md (approved, CLEAN, appends lesson 20). This one inserts inside lesson 13's body at lines 211–236 and renumbers nothing. Disjoint regions, so the two merge cleanly in either order — I checked the numbering specifically, since a second lesson-adding PR is exactly where a duplicate ordinal would hide, and this one adds no ordinal at all.

One observation, not a blocker. The addition prescribes "link the source, or write unverified next to it", and it complies — but it also demonstrates a third state the rule does not name: a source that is cited, checked, and still only evidence, because the page says its own data is incomplete. The PR handles this correctly in the prose (that is precisely why it lands on @supports rather than a version comparison), yet the rule as stated offers only two labels. A reader applying the rule mechanically would mark that citation "verified" and move on. Worth a clause; not worth holding the PR.

Scope: I read the full diff (25 added lines, one file), fetched and checked the cited page myself, and confirmed CI at this head (17 CheckRuns SUCCESS, license/cla SUCCESS). I did not verify the Safari 16.2 threshold — which the PR already marks unverified, so we are in the same state on it, and saying so is the point.

Base is BEHIND; yours to update, and it will stale this approval.

[github-actions]

@cla-assistant check

[sonichi]

Reviewed at 0c45c16a (COMMENT — owner-login account). Docs-only: one clause appended to REVIEW.md lesson 13, no new lesson number, and the clause is the reviewer-side mirror of what lesson 13 already says about formatters — cite or label every external fact in the review's own prose. Checked the grounding instance's claim shape against the cited Tauri table link (present, and the clause itself labels the 16.2 threshold unverified, which is the rule applied to itself). Nothing here changes a mechanism; it changes what a review must carry. No findings; I recommend merge when the owner authorizes it (2 write approvals stand: qingyun-wu, yixuan-ag2).

Stand: Echo Act IV Mini