comments: converge two repeated blocks — and the ± note was asserting the opposite of the truth

[christopherthompson81]

You asked for consistency over letter-by-letter, so this is the one uniform pass: the ± note (25 files) and the NOT QUITE VERBATIM inventory note (5 files), each of which had drifted into several wordings.

Converging them turned up a defect the duplication had been hiding.

The ± note was asserting the opposite of the truth

Every copy asserted an ordering — "Runs BEFORE the + rule" — and then gave as its reason the fact that makes that ordering irrelevant: "± is a single character, so the + rule cannot see it." Both halves cannot be true.

The second one is. ± is U+00B1; + is U+002B:

/(^|[\s(])\+\s?(\d)/u.test("±5")   →   false

Kazakh is the proof by counterexample. Its ± rule runs after two + rules and has always been correct — which is exactly what exposed this: I wrote a check to verify the claim across all 21 files rather than trust it, and Kazakh failed. (Luxembourgish also flagged, but that was my detector not recognising named-constant regexes.)

So the ordering is free, and the note now says the thing that actually matters:

⚠ ± IS A SINGLE CHARACTER (U+00B1), NOT A +, so no + rule can ever match inside it. It needs its own rule or the sign is dropped in silence; ordering against the + rule is free.

That is the useful warning — an unclaimed sign is dropped, not misread, so it is invisible to every check that looks for a wrong reading.

Also dropped

• "zero new sourcing" / "nothing is invented" — justification aimed at a reviewer of the original change, not at a porter reading the file;
• "the FORM is the one every language that already read ± uses (bg/da/is/nb/ro/sv…)" — a survey of other languages, in files about Macedonian, Welsh, Japanese…;
• "± IS NOW FREE, and it was not before this commit" (hu, pl) — a dangling reference to a commit nobody reading the file can find.

Kept — and deliberately not propagated

Four things were load-bearing but present in only some copies. Copying them everywhere would have been inventing claims about languages that never made them, so each stays exactly where it was:

kept	where
SIGN names vs OPERATION names — ± marks a tolerance, not an addition	af, ca, kk, cy
the same distinction, with Korean's own pair: 더하기/빼기 are operators, 플러스/마이너스 are signs	ko
min comes from the register rather than from another rule, because this language has no true minus sign	nl
wiki citations naming the sign word against the glyph	tr, hu

NOT QUITE VERBATIM

Lost four lines of history and an anecdote about a Romanian word (Thérèse → ˈthrese) that appeared in files for five other languages. What remains is the fact: these letters are absent because the g2p has no rule for them, and native-inventory.test.ts measures that rather than trusting it.

Verification

Comment-only — emitted code byte-identical across all 32 files (--removeComments, both revisions, diff -r). Typecheck clean, 3118 tests pass.

One code finding, not fixed here

polish/normalize.ts runs s = s.replace(/±/gu, " plus minus ") twice, at lines 367 and 373, each with its own comment block. The second is dead — no ± survives the first. Pre-existing on main; a separate one-line PR follows, so this one keeps its byte-identity guarantee.