@redocly/recheck 0.1.0 → 0.3.0

package/presets/google/PROVENANCE.md

@@ -0,0 +1,1022 @@
+# Provenance: `recheck/google`
+
+Source: [Google developer documentation style guide](https://developers.google.com/style)
+(canonical URL: `https://developers.google.com/style`). License:
+[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Sync date: **2026-07-29**.
+
+Modification note: rules are adapted to Recheck's assertion vocabulary
+(`swap`, `pattern`, `capitalization`, `length`, plus a handful of
+markdownlint-parity/Recheck-original token rules); wording is paraphrased
+into each rule's `message`, not quoted verbatim from the guide. Every rule
+carries a `link:` to its source page.
+
+## How this table was produced
+
+Five independent verification passes fetched the live guide directly
+(`curl`, not a summarizing fetch) and confirmed or rejected each candidate
+rule against the raw page text/HTML: four slice verifiers (`task-9-verify-A.md`
+covering style guide §2.1-2.5, `task-9-verify-B.md` §2.6-2.10,
+`task-9-verify-C.md` §3.1-3.5, `task-9-verify-D.md` §3.6-3.9) plus one
+cross-check pass that re-parsed the word-list page with a stricter HTML5
+parser and diffed every replacement value quoted by the other four
+(`task-9-verify-crosscheck.md`). Together they checked ~380 candidate
+rules/entries and found **6 fabrications** — rules that appeared in an
+earlier research draft but do not exist anywhere on the live guide (see
+"Fabrications" below). This preset is built **only** from entries those
+five reports marked `CONFIRMED`; nothing here was sourced from the research
+draft directly. Fetch date for the verification passes and for
+`sources.json`'s hashes is the same day this preset was authored: 2026-07-29.
+
+Every quote below is the verifier's own quote (or fetch-log citation),
+reproduced here at second hand — see the verifier reports themselves for
+the full text and additional context.
+
+## Shipped rules
+
+Severity policy: `error` is reserved for rules checking pure document
+STRUCTURE (heading hierarchy/uniqueness, list-item mechanics, table
+mechanics, link placement, alt-text presence, sentence length) where a
+violation is unambiguous and mechanical; every word-choice, terminology,
+punctuation-convention, and phrasing rule is `warn`. This is a
+simplification of spec §2's severity table (which gives a shorter, purely
+illustrative example list) applied uniformly here for predictability, at
+the cost of a few punctuation-mechanics rules (Oxford comma, en dash,
+single-space-between-sentences) that could arguably also be `error` — see
+"Author's judgment calls" at the end of this document.
+
+Four shipped rules are named exceptions to that split, not silent
+inconsistencies with it: `no-code-in-heading` is heading-scoped (in the
+STRUCTURE family above by location) but ships at `warn` because the guide
+states it with hedged wording ("Avoid code items in headings," not
+"don't"). `no-numbered-headings` is also heading-scoped and the guide
+states it unconditionally ("Don't use numbers in headings..."), but ships
+at `warn` because the shipped pattern is a narrowed heuristic (bare
+leading ordinals and `Step N`/`Part N` markers only, to keep the
+false-positive rate low), not a complete detector of every way a heading
+could number a sequence — the rule's confidence is in what it does flag,
+not full coverage of the guide's stated principle. `emphasis-style` and
+`strong-style` ship at `warn` because the guide states its markup
+preference as a recommendation ("we recommend underscores," "it's best to
+use double asterisk"), not an unconditional "don't." So the actual policy
+is: STRUCTURE, stated by the guide as an unconditional, completely
+detectable rule → `error`; everything else — including these four
+nominally-structural rules the guide itself hedges, or that need an
+intentionally incomplete detection heuristic → `warn`.
+
+### Structural (heading, list, table, link, alt-text, sentence mechanics) — `error`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/heading-sentence-case` | [headings](https://developers.google.com/style/headings) | "Use sentence case for all headings and titles." | CONFIRMED |
+| `google/heading-increment` | [headings](https://developers.google.com/style/headings) | "put an `<h3>` tag only under an `<h2>` tag" | CONFIRMED |
+| `google/single-h1` | [headings](https://developers.google.com/style/headings) | "only use a level-1 heading once on a page" | CONFIRMED |
+| `google/first-line-h1` | [headings](https://developers.google.com/style/headings) | "only use a level-1 heading once on a page" (same statement as `single-h1`; see "Author's judgment calls") | CONFIRMED |
+| `google/no-duplicate-heading` | [headings](https://developers.google.com/style/headings) | "easier to jump between pages and sections...if the headings...are unique" | CONFIRMED |
+| `google/no-trailing-punctuation` | [periods](https://developers.google.com/style/periods) | "Don't end headings with periods." | CONFIRMED |
+| `google/no-empty-headings` | [headings](https://developers.google.com/style/headings) | "Don't use empty headings. Make sure headings are followed by content." | CONFIRMED |
+| `google/no-emphasis-as-heading` | [accessibility](https://developers.google.com/style/accessibility) | "Tag headings using heading elements." | CONFIRMED (nuance: the engine's own token rule excludes a bold/italic lead-in followed by more text in the SAME paragraph, and a bold/italic-only paragraph that ends in required punctuation — but NOT a bold/italic-only paragraph with no ending punctuation whose description follows in a later paragraph, which the guide's own "run-in heading" pattern also permits; see verifier A row 12 and "Known limitations" below) |
+| `google/no-link-in-heading` | [headings](https://developers.google.com/style/headings) | "Don't put links in headings." | CONFIRMED |
+| `google/list-item-capital` | [lists](https://developers.google.com/style/lists) | "Start each list item with a capital letter" | CONFIRMED |
+| `google/no-alt-text` | [accessibility](https://developers.google.com/style/accessibility) | "For every image, provide an alt attribute" | CONFIRMED |
+| `google/no-merged-cells` | [accessibility](https://developers.google.com/style/accessibility) (also [tables](https://developers.google.com/style/tables)) | "Don't merge cells. Don't use colspan or rowspan attributes" | CONFIRMED |
+| `google/sentence-length` | [accessibility](https://developers.google.com/style/accessibility) | "Try to use fewer than 26 words per sentence." | CONFIRMED — mapped to the new `length` assertion (`unit: words`, `max: 25`) per spec §5.6's stated-numbers table; this is the first non-prose preset to ship a `length`-backed rule (see "Engine/registry changes" below) |
+
+### Headings (residual) / lists — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/no-code-in-heading` | [headings](https://developers.google.com/style/headings) | "Avoid code items in headings." | CONFIRMED |
+| `google/no-numbered-headings` | [headings](https://developers.google.com/style/headings) | "Don't use numbers in headings to indicate a sequence" | CONFIRMED |
+| `google/list-length` | [lists](https://developers.google.com/style/lists) | "a single item isn't really a list" | CONFIRMED, but the verifier marked this specific line NOT-ENFORCEABLE (descriptive aside, not an imperative rule) — downgraded from the generic "list mechanics" `error` class to `warn` for that reason; ships via `list-length`'s own `min: 2` default, no `max` (Google states no upper bound; the 2-7 range is Microsoft's, spec §5.6) |
+
+### Voice, person, tense, contractions — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/second-person` | [person](https://developers.google.com/style/person) | "use `you` or `your` instead of `we`, `our`, or `us`" | CONFIRMED (organizational-reference exception noted in the rule's message; detection-only). Fix wave A: `ignoreCase: true` replaced with an explicit `We/we/Our/our/Us/us` alternation so it stops also matching the all-caps abbreviation "US" — see "Fix wave A corrections" |
+| `google/use-contractions` | [contractions](https://developers.google.com/style/contractions) | "we recommend using negation contractions such as isn't, don't, and can't" | CONFIRMED (`fix: false`: guide's own "is *not*" emphasis exception) |
+| `google/no-triple-contractions` | [contractions](https://developers.google.com/style/contractions) | "Don't use three-word contractions such as mightn't've." | CONFIRMED |
+| `google/no-lets` | [word-list#lets](https://developers.google.com/style/word-list#lets) | "Don't use if at all possible." | CONFIRMED — citation corrected: the draft cited "contractions", but `let's` is never mentioned there; the quote lives on the word-list page (verifier A row 14) |
+| `google/no-please-note` | [word-list#please](https://developers.google.com/style/word-list#please) | "Don't use please in the normal course of explaining how to use a product" (the phrase `please note` specifically has no documented exception) | CONFIRMED. Fix wave A: `fix: false` added — the delete-swap left a capitalization/fragment mess behind ("Please note that X." -> "that X."); see "Fix wave A corrections" |
+| `google/no-please` | [word-list#please](https://developers.google.com/style/word-list#please) | "Use please only when you're asking for permission or forgiveness...Recommended: If the issue persists, please contact your account representative." | CONFIRMED, **must be DETECT-ONLY, never a swap** — the guide's own recommended example sentence uses "please"; a delete-swap would rewrite text the guide endorses (per task-9-author-corrections.md) |
+
+### Timeless documentation — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/no-timeless-phrases` | [timeless-documentation](https://developers.google.com/style/timeless-documentation) | "as of this writing, currently, does not yet, eventually, existing, future...latest, new, newer, now, old, older, presently, at present, soon" (full confirmed list) | CONFIRMED, but only 5 of the ~15 confirmed terms are shipped (`as of this writing`, `at present`, `presently`, `does not yet`, `currently`) — the rest are ordinary high-frequency words (`existing`, `future`, `latest`, `new`, `newer`, `now`, `old`, `older`, `soon`, `eventually`, `in the future`) with legitimate everyday uses that would make a blind pattern unusably noisy; see "Author's judgment calls" |
+
+### Latinisms, abbreviations, slang — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/no-latinisms` | [word-list](https://developers.google.com/style/word-list) | "Don't use i.e. or e.g."; "Don't use vs. as an abbreviation for versus" | CONFIRMED (i.e./e.g./vs. — independently reconfirmed in verifier C's 3.1 slice). Fix wave A: carries only the 3 period-terminated keys now (see "Fix wave A corrections" below for why `aka`/`vice versa` moved out) |
+| `google/no-latinisms-plain` | [word-list](https://developers.google.com/style/word-list) | "aka: Don't use. Instead, write out also known as"; "vice versa: ...use...the other way around" | CONFIRMED, same verification as `google/no-latinisms` above. Split out in Fix wave A (2026-07-29) so these two non-period keys get full `\b...\b` anchoring instead of the period-terminated group's unanchored match — see "Fix wave A corrections" |
+| `google/no-internet-slang` | [abbreviations](https://developers.google.com/style/abbreviations) | "Don't use internet slang abbreviations such as tl;dr, ymmv, RTFM" | CONFIRMED (each also has its own word-list replacement: "To summarize" / "Your results might vary" / "For more information, see...") |
+| `google/no-via` | [word-list#via](https://developers.google.com/style/word-list#via) | "Don't use." | CONFIRMED, DETECT-ONLY (no replacement given) — this is the exact term the research file's own provenance note flagged a summarizer once inverted; independently reconfirmed by verifier C and the cross-check |
+| `google/abbrev-no-periods` | [abbreviations](https://developers.google.com/style/abbreviations) | "Don't use periods with acronyms or initialisms." | CONFIRMED |
+| `google/us-abbreviation` | [word-list#US](https://developers.google.com/style/word-list#US) | "OK to use as an abbreviation for United States. Don't use U.S. or U.S.A." | CONFIRMED |
+| `google/no-slash-abbrev` | [slashes](https://developers.google.com/style/slashes) | "Slashes with abbreviations... Recommended: care of, with / Not recommended: c/o, w/" | CONFIRMED. Fix wave A: both keys gained a leading-only `\b` (were matching inside "src/output", "www/static", "show/hide", "new/old"); Fix wave C: both also gained a trailing `(?![A-Za-z])` (were matching inside "w/o", "c/oscillator") — see "Fix wave A corrections" / "Fix wave C corrections" |
+
+### Numbers, dates, units — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/spell-out-ordinals` | [numbers](https://developers.google.com/style/numbers) | "Not recommended: 1st, 5th, 12th, 43rd" | CONFIRMED |
+| `google/number-format` | [numbers](https://developers.google.com/style/numbers) (also [hyphens](https://developers.google.com/style/hyphens) for the "from N-M" token) | "use numerals and the percent sign (%), without a space"; "place a zero in front of the decimal point"; "Recommended: 192x192 / Not recommended: 192 x 192"; "Not recommended: from 8-20 files" | CONFIRMED |
+| `google/date-format` | [dates-times](https://developers.google.com/style/dates-times) | "Not recommended: 12/02/2017" | CONFIRMED |
+| `google/time-format` | [word-list#AM,\_PM](https://developers.google.com/style/word-list#AM,_PM) | "use all caps, no periods, and a space before" (AM/PM); "Remove the minutes from round hours" (dates-times page) | CONFIRMED — citation corrected: the AM/PM quote lives on the word-list page, not dates-times (verifier B row 63) |
+| `google/rfc-spacing` | [word-list#RFC](https://developers.google.com/style/word-list#RFC) | "use a space between RFC and the number (for example, RFC 2318)" | CONFIRMED |
+| `google/data-rate-units` | [word-list#GBps](https://developers.google.com/style/word-list#GBps) | "By convention, we don't use KB/s" (and the other 5 unit pairs) | CONFIRMED — the master research table's "8 pairs" claim was wrong; verified exactly 6 (verifier B row 67, cross-check) |
+
+### Punctuation — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/no-ampersand` | [text-formatting](https://developers.google.com/style/text-formatting) | "Don't use ampersands (&) as conjunctions or shorthand for and." | CONFIRMED (UI-element exception preserved: pattern requires spaces on both sides, so `AT&T`/`&amp;` are untouched) |
+| `google/dash-style` | [dashes](https://developers.google.com/style/dashes) | "En dashes...Don't use. Instead, use a hyphen or the word to."; "Don't put a space before or after it [the em dash]." | CONFIRMED |
+| `google/single-space-sentences` | [periods](https://developers.google.com/style/periods) | "Leave only one space between sentences." | CONFIRMED |
+| `google/conjunctive-adverb-comma` | [commas](https://developers.google.com/style/commas) | "put a comma after the conjunctive adverb" | CONFIRMED — the draft's fourth example ("Nonetheless") is not named on the live page; only the three confirmed examples (otherwise/however/therefore) are shipped |
+| `google/comma-before-that` | [pronouns](https://developers.google.com/style/pronouns) | "That introduces a restrictive clause. It isn't preceded by a comma." | CONFIRMED |
+| `google/neither-nor` | [word-list#neither](https://developers.google.com/style/word-list#neither) | "Write neither A nor B, not neither A or B." | CONFIRMED |
+| `google/no-and-or` | [slashes](https://developers.google.com/style/slashes) | "avoid writing and/or except when space is limited, such as in tables" | CONFIRMED — the table exception is not separately carved out in the rule's scope (see "Known simplifications") |
+
+### Links — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/vague-link-text` | [cross-references](https://developers.google.com/style/cross-references) | "Don't use phrases such as this document, this article, or click here" | CONFIRMED (page renamed from `link-text`; content unchanged) |
+| `google/no-url-as-link-text` | [cross-references](https://developers.google.com/style/cross-references) | "don't use a URL as link text" | CONFIRMED (legal/ToS exception noted, not separately modeled — see "Known limitations" below) |
+| `google/link-intro-about` | [cross-references](https://developers.google.com/style/cross-references) | "Don't use on instead of about" | CONFIRMED |
+| `google/link-punctuation` | [cross-references](https://developers.google.com/style/cross-references) | "don't put the link text in quotation marks" | CONFIRMED |
+| `google/no-target-blank` | [cross-references](https://developers.google.com/style/cross-references) | "Don't force links to open in a new tab or window" | CONFIRMED |
+| `google/self-reference-terms` | [word-list#documentation](https://developers.google.com/style/word-list#documentation) | "use this document, and not this article, this topic, this doc" | CONFIRMED (general terminology preference, independent of link text — distinct from `vague-link-text` above, which is scoped to the `link` segment) |
+
+### Text formatting — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/emphasis-style` | [text-formatting](https://developers.google.com/style/text-formatting) | "we recommend underscores" | CONFIRMED |
+| `google/strong-style` | [text-formatting](https://developers.google.com/style/text-formatting) | "best to use the double asterisk for bold" | CONFIRMED |
+| `google/no-underline` | [text-formatting](https://developers.google.com/style/text-formatting) | "Reserve underlining for link text" | CONFIRMED |
+| `google/no-casing-style-names` | [capitalization](https://developers.google.com/style/capitalization) | "Don't use a casing style name, such as camel case or snake case" | CONFIRMED |
+
+### Code in text — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/no-inflected-code` | [code-in-text](https://developers.google.com/style/code-in-text) | "Don't inflect the name of a code element" | CONFIRMED |
+
+### UI elements and verbs — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/ui-element-quotes` | [ui-elements](https://developers.google.com/style/ui-elements) | 'click the "Next" button' (Not recommended example) | CONFIRMED |
+| `google/no-click-on` | [word-list#click](https://developers.google.com/style/word-list#click) | "Don't use click on." | CONFIRMED |
+| `google/no-hover` | [word-list#hover](https://developers.google.com/style/word-list#hover) | "Don't use. Instead use hold the pointer over." | CONFIRMED — DETECT-ONLY, not a swap (inflected forms "hovering"/"hovered" don't slot into the replacement phrase) |
+| `google/no-uncheck` | [word-list#uncheck](https://developers.google.com/style/word-list#uncheck) | "use clear for checkboxes" | CONFIRMED. Bare `check` and `deselect` are deliberately NOT shipped — see "Author's judgment calls" |
+| `google/scroll-to` | [word-list#scroll](https://developers.google.com/style/word-list#scroll) | "write go to the section, instead of scroll to the section" | CONFIRMED (preference, not a flat ban — `fix: false`) |
+| `google/no-toggle-verb` | [ui-elements](https://developers.google.com/style/ui-elements) | "Don't use the word toggle as a verb. Describe the action." | CONFIRMED — citation correction: not a word-list entry as the draft implied, but is on the `ui-elements` page (verifier D) |
+| `google/keyboard-keys` | [ui-elements](https://developers.google.com/style/ui-elements) | "Spell out the names of modifier keys"; "use uppercase instead of lowercase" | CONFIRMED |
+| `google/chapter-terminology` | [word-list#chapter](https://developers.google.com/style/word-list#chapter) | "Instead, refer to documents, pages, or sections" | CONFIRMED, DETECT-ONLY |
+
+### Plain language and wordiness — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/plain-language-swaps` | [word-list](https://developers.google.com/style/word-list) | "allows_you_to: Don't use. Instead, use lets you"; "enable: Not recommended: ...enables you to..."; "comprise: Don't use. Instead, use consist of..."; "desire: Don't use. Instead, use a word like want or need"; "learnings: Don't use. Instead, refer to knowledge..."; "agnostic: Don't use. Instead, use...platform-independent" | CONFIRMED |
+| `google/in-order-to` | [word-list#in_order_to](https://developers.google.com/style/word-list#in_order_to) | "Avoid in order to; instead, use to. Use in order to when needed to clarify meaning." | CONFIRMED (conditional — `fix: false`) |
+| `google/utilize` | [word-list#utilize](https://developers.google.com/style/word-list#utilize) | "Use with caution. Don't use utilize when you mean use. It's OK to use utilize... when referring to the quantity of a resource being used." | CONFIRMED, DETECT-ONLY given the documented exception |
+| `google/leverage` | [word-list#leverage](https://developers.google.com/style/word-list#leverage) | "Avoid using if you mean use... use, build on, or take advantage of." | CONFIRMED (`fix: false` — three valid alternatives given) |
+| `google/performant` | [word-list#performant](https://developers.google.com/style/word-list#performant) | "Avoid where possible. Instead, use a more precise term." | CONFIRMED, DETECT-ONLY (no fixed replacement) |
+| `google/copy-and-paste` | [word-list#Copy\_and\_paste](https://developers.google.com/style/word-list#Copy_and_paste) | "Avoid using. Instead, explain what to enter into a field and not how." | CONFIRMED |
+| `google/create-a-new` | [word-list#Create\_a\_new](https://developers.google.com/style/word-list#Create_a_new) | "Avoid using unless you need to distinguish... Instead, use Create a ..." | CONFIRMED (exception noted; `fix: false`) |
+| `google/no-run-the-following-command` | [procedures](https://developers.google.com/style/procedures) | "Avoid using run the following command to introduce code. Instead, focus on what the command does." | CONFIRMED |
+| `google/cons-and-pros` | [word-list#pros](https://developers.google.com/style/word-list#pros) | "cons: Don't use. Instead, use a more precise term, such as disadvantages." / "pros: ...such as advantages." | CONFIRMED — only the compound phrase "pros and cons" ships; bare "pros"/"cons" are excluded, see "Author's judgment calls" |
+
+### Product and brand names — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/product-names` | [word-list](https://developers.google.com/style/word-list) | "Cloud SDK: Not Google Cloud SDK"; "APIs Explorer: Not API explorer..."; "API key: Not developer key or dev key"; "account name: ...use username"; "curated roles: ...use predefined roles"; "network IP address: ...use internal IP address"; "media type: ...Don't use MIME type"; "curl: Not cURL"; "interconnect type: ...use connection type"; "peering zone: Not peer zone"; "Android-powered device: Not Android device"; "Not...Cloud Platform, or Cloud" (Google Cloud); "Cloud console: Not...Developers Console" | CONFIRMED (each entry individually confirmed in verifier D's §3.4). Fix wave A: `'Cloud console'` gained a `(?<!Google )` lookbehind (self-compounding fix — see "Fix wave A corrections"); `GCP` moved to `google/gcp-name` below. Fix wave C: the lookbehind widened to `(?<![Gg][Oo][Oo][Gg][Ll][Ee]\s+)` (case-insensitive, whitespace-tolerant — see "Fix wave C corrections") |
+| `google/gcp-name` | [word-list](https://developers.google.com/style/word-list) | "Not GCP...(Google Cloud)" | CONFIRMED, same verification as `google/product-names` above. Split out in Fix wave A (2026-07-29), `fix: false` — see "Fix wave A corrections". Fix wave C: `applyMatchCase` fixed at the engine, `fix: false` REMOVED — see "Fix wave C corrections" |
+| `google/brand-capitalization` | [word-list](https://developers.google.com/style/word-list) | "Google Play services: Write services in lowercase."; "Google Account, Google Accounts: Capitalize Account."; "Markdown: Always capitalized."; "Material Design: Capitalize each word."; "Search Console: Capitalize each word." | CONFIRMED |
+
+### Compound and one-word forms — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/compound-forms` | [word-list](https://developers.google.com/style/word-list) (webpage via [hyphens](https://developers.google.com/style/hyphens)) | ~68 individual "Not X" / "X: Not Y" entries, e.g. "email: Not e-mail, Email, or E-mail."; "checkbox: Not check box."; "frontend: Not front-end or front end." | CONFIRMED (verifier C's 3.3 slice: ~68/70 confirmed; see "Fabrications"/"Dropped as unverified" for the 2 that didn't make it). Fix wave A: `colo` moved to `google/colo-form` below; `'in line': 'inline'` dropped (see "Dropped in Fix wave A" above) |
+| `google/colo-form` | [word-list](https://developers.google.com/style/word-list) | "colocate: ...Not co-locate or colo." | CONFIRMED, same verification as `google/compound-forms` above. Split out in Fix wave A (2026-07-29), `fix: false` (noun/verb mismatch) — see "Fix wave A corrections" |
+| `google/acronym-forms` | [word-list](https://developers.google.com/style/word-list) | "HTTPS: Not HTTPs."; "IPsec: Not IPSec."; "NoSQL: Not No-SQL or No SQL."; "OAuth 2.0: Not OAuth 2, OAuth2, or Oauth."; "microservices: Not micro-services."; "fintech: ...Don't use FinTech or fin-tech."; "ad tech: ...Don't use adtech or ad-tech." | CONFIRMED. `I-O`/`IO` → `I/O` confirmed individually. Fix wave A: `UNICODE`/`IPSEC` moved to `google/acronym-caps-detect-only` below; `SHA1` moved to `google/sha1-form` below; `Microservices` (capitalized form) dropped (see "Dropped in Fix wave A" above) |
+| `google/acronym-caps-detect-only` | [word-list](https://developers.google.com/style/word-list) | "Unicode: Not UNICODE."; "IPsec: Not...IPSEC." | CONFIRMED, same verification as `google/acronym-forms` above. Split out in Fix wave A (2026-07-29), `fix: false` (case-preservation round-trip made the fix a permanent no-op) — see "Fix wave A corrections" |
+| `google/sha1-form` | [word-list](https://developers.google.com/style/word-list) | "SHA-1: Not SHA1, except in string literals or enum values, and in hyphenated phrases such as HMAC-SHA1." | CONFIRMED (verifier C row 202). Split out in Fix wave A (2026-07-29), `fix: false` per this file's own header policy for rules with a documented guide exception — see "Fix wave A corrections" |
+
+### Inclusive language / ableist language / jargon with people references — `warn`
+
+| Rule id | Source URL | Quote | Verdict |
+|---|---|---|---|
+| `google/master-slave` | [word-list#slave](https://developers.google.com/style/word-list#slave) | "Don't use. Instead, use alternative terms... such as worker or replica." | CONFIRMED. Only "slave" ships; bare "master" is deliberately excluded — see "Author's judgment calls" |
+| `google/blacklist-whitelist` | [word-list#blacklist](https://developers.google.com/style/word-list#blacklist) | "For the noun blacklist, consider... denylist, excludelist, or blocklist" / "whitelist...consider...allowlist, trustlist, or safelist" | CONFIRMED. Noun forms only — the guide itself says a word-for-word swap isn't the best fix for verb forms |
+| `google/black-white-hat` | [word-list#blackhat](https://developers.google.com/style/word-list#blackhat) | "Don't use. Instead, use precise terms... such as illegal, unethical, or in violation of rules." | CONFIRMED |
+| `google/black-white-box-testing` | [word-list#black-box](https://developers.google.com/style/word-list#black-box) | "For monitoring, use synthetic monitoring. For testing, use opaque-box testing." | CONFIRMED |
+| `google/grayed-out` | [word-list#grayed-out](https://developers.google.com/style/word-list#grayed-out) | "Don't use. Instead, use unavailable." | CONFIRMED |
+| `google/grandfathered` | [word-list#grandfathered](https://developers.google.com/style/word-list#grandfathered) | "Instead, use an adjective like legacy or exempt or a verb like made an exception." | CONFIRMED |
+| `google/gendered-terms` | [word-list#man\_hours](https://developers.google.com/style/word-list#man_hours) | "man hours...Instead use terms like person hours"; "male adapter...use plug"; "he/she...use...they" | CONFIRMED |
+| `google/jargon-with-people-references` | [word-list#ninja](https://developers.google.com/style/word-list#ninja) | "ninja: Don't use to refer to a person. Instead, use a term such as expert."; "DMZ...use...perimeter network"; and others | CONFIRMED |
+| `google/ableist-figurative-terms` | [word-list#crazy](https://developers.google.com/style/word-list#crazy) | "use complicated, complex, baffling, strange, or unexpected...only for inanimate objects" | CONFIRMED. Restricted to the figurative/inanimate-object sense; "mad" and "hang"/"hung" excluded (see "Author's judgment calls") |
+| `google/dummy-variable` | [word-list#dummy-variable](https://developers.google.com/style/word-list#dummy-variable) | "Don't use to refer to placeholders. Instead, use placeholder." | CONFIRMED |
+| `google/blind-figurative` | [word-list#blind](https://developers.google.com/style/word-list#blind) | "blind to, blind eye to...use more precise terms like ignore, unaware of, disregard, avoid, or reject" | CONFIRMED, restricted to the figurative sense only (the SAME entry also covers "blind writes"/"blind change", real technical terms, deliberately not matched) |
+| `google/unsighted-visually-challenged` | [word-list#unsighted](https://developers.google.com/style/word-list#unsighted) | "Don't use. See blind." | CONFIRMED — resolved to the PERSON-REFERENCE sense of "blind" (person who is blind/visually impaired/low-vision), NOT the figurative sense above, per task-9-author-corrections.md's explicit correction |
+| `google/disability-language` | [inclusive-documentation](https://developers.google.com/style/inclusive-documentation) | "avoid terms such as the disabled or a quadriplegic"; "such as victim of, suffering from... instead, use... experiencing, living with" | CONFIRMED |
+| `google/technical-jargon-precision` | [word-list#fat](https://developers.google.com/style/word-list#fat) | "use a precise modifier... high-capacity network connection instead of fat connection" | CONFIRMED — framed as technical-jargon precision, NOT ableist language, per task-9-author-corrections.md: Google's own entries for "chubby"/"fat" never mention people |
+
+## Excluded candidates
+
+Every candidate below was checked by one of the five verification passes
+(or, where marked, judged independently by this preset's author) and is
+**not** shipped, with the reason. This section exists so "why doesn't
+`recheck/google` check X?" has a documented answer instead of looking like
+an oversight.
+
+### Fabrications (found nowhere on the live guide) — never ship
+
+| Candidate | Why excluded |
+|---|---|
+| `flesch-reading-ease` metric threshold | No basis anywhere in the guide — zero hits for flesch/reading-ease/readability/grade-level across every fetched page. The guide states only the qualitative principle ("use simpler words and shorter sentences"). Spec §5.6 and spec:215 independently rule out a `metric` rule in the flagship presets; shipping this would attribute an invented numeric mandate to Google under a CC-BY citation. |
+| `respective`/`respectively` → rewrite | No such headword or rule exists anywhere in the word list or any topic page fetched (verifier C, independently re-confirmed by the cross-check). |
+| `makes use of` → `uses` | Zero occurrences anywhere in the fetched pages (verifier C, cross-check). |
+| `query string (not querystring)` | No entry anywhere on the word-list page or any cited topic page; a targeted search of the guide also returns nothing (verifier C, cross-check). |
+| `real time`/`real-time` noun/verb pair | Zero hits in the parsed text or the raw HTML (verifier B, cross-check). |
+| `robust` as a caution term | Does not appear anywhere in the current word list (verifier B, cross-check). |
+
+### Dropped as unverified (real principle, no dedicated confirmable entry)
+
+| Candidate | Why excluded |
+|---|---|
+| `user name` → `username` | No dedicated word-list entry; only inferable from `account name → username` and the general closed-compound principle. Per the verification contract, unverified means dropped, not shipped at a lower confidence. |
+| N/A first-reference spell-out | The word-list entry only requires spelling out `N/A`/`NA` on FIRST reference, not a blanket ban — Recheck has no reliable way to track "is this the first reference" positionally across a document, so it's excluded rather than shipped as a (wrong) blanket swap. |
+
+### TOO-RISKY (guide confirms it, but the "avoid" token is too ordinary/polysemous to blind-match)
+
+These are all real, guide-stated preferences; none are shipped because the
+literal string is common enough in unrelated, correct usage that a blind
+`swap`/`pattern` would generate far more false positives than true ones.
+
+| Candidate | Why excluded |
+|---|---|
+| `abort`, `kill`, `terminate` → stop/exit/cancel/end | Standard, correct technical vocabulary (process signals, transaction aborts, connection termination) with a guide-documented command-line-syntax carve-out a blind swap can't detect. |
+| `execute` → run; `access` (verb); `possible`/`impossible`; `impact` (verb); `interface` (verb); `exploit`; `scale` (bare); `review` (="first read"); `each` (="all") | Each has a real but conditional/sense-scoped ruling; the bare word is common enough in the OTHER (unflagged) sense that a blind match would misfire constantly. |
+| `simply`/`simple`, `easy`/`easily`, `quick`/`quickly`, `just` | The guide's own "try eliminating this word" framing has documented legitimate uses ("just" is explicitly OK "to convey that one approach is simpler"); a blind delete-swap fights the guide's own exception. |
+| `America`/`American` (=USA sense) | Scoped to the USA-sense only; "American" (American English, American Express, Latin American) has far more legitimate uses than violations. |
+| `Cloud` (bare, capitalized, standing for Google Cloud); `portal`/`dashboard` (meaning the Google Cloud console) | The guide explicitly permits lowercase "the cloud" generically, and "portal"/"dashboard" are ordinary words in enormously common non-Google-Cloud technical writing. |
+| `PostgreSQL`/`Postgres`, `directory`/`folder`, `plain text`/`plaintext` | Context-conditional (match the UI's own wording; CLI vs. GUI; cryptography context only), not a free either-or `consistency` pair — shipping as unconditional would mis-enforce legitimate context-appropriate variation. |
+| `firewalls` → `firewall rules` | Scoped to Compute Engine/networking documentation only; "outside of [that], the term firewalls is acceptable" per the guide's own text — Recheck can't detect document subject matter. |
+| `deselect` (bare, banned) | **Wrong to ship at all**: `deselect` is Google's own correct/recommended term for NON-checkbox UI elements ("use clear for checkboxes, and deselect for other UI elements"). Only `uncheck` (unambiguously checkbox-only in meaning) is shipped. |
+| bare `check` (verb) | Extremely polysemous ("check the logs", "check that X is true") — only the unambiguous `uncheck` ships. |
+| `hit` (UI click synonym) | "hit" is extremely common in unrelated technical senses (cache hit, rate limit hit); the guide's UI-click sense can't be isolated by a blind pattern. |
+| `type` → `enter` | "type" is one of the most common words in technical prose in unrelated senses (data type, type of X); too ambiguous to match blindly. |
+| `menu item`/`choice`/`option` → `command` | "option"/"choice" are common ordinary words far outside the menu-item sense the guide targets. |
+| bare `master` | Google's own quote scopes the objection to the master/slave PAIRING ("Never use in conjunction with slave"), not the standalone word, which has many unrelated legitimate senses (master's degree, master key, master bedroom). Only `slave` ships. |
+| `drag and drop` (as opposed to `click and drag`) | Commonly and correctly used as a noun/adjective ("drag-and-drop interface"); only the unambiguous verb phrase `click and drag` ships. |
+| `hang`/`hung` (of a system) | Extremely polysemous outside the system sense ("hung the picture", "hung jury", "hang up the phone"). |
+| `healthy` (of a system) → responsive | Polysemous even in technical prose ("a healthy amount of caution", "healthy competition"). |
+| `mad` (ableist figurative sense) | At least as commonly used to mean "angry", a sense the guide never objects to. |
+| `abnormal`, `deficient`, `deformed` (of a person) | Explicitly "OK to refer to a condition of a computer system"; unscoped, these would misfire on ordinary error-handling prose ("abnormal termination", "abnormal exit code"). |
+| `gimp`/`gimpy`, `lame` | `gimp` has an explicit carve-out for the GIMP image editor and similarly-named tools; neither has a fixed replacement token, and both are DETECT-ONLY at best. |
+| `native` (of people); `target` (verb, of people) | Both are advisory ("avoid... when possible") and the words themselves are ubiquitous in unrelated, correct senses ("native app", "cloud-native", "target audience"). |
+| `above`/`below`/`higher`/`lower`/`older` (version-range words) | Explicitly OK when non-directional ("below average", describing a hierarchy) and reversed for Android docs; only safe when anchored to an unambiguous version-number context, which a blind pattern can't establish. |
+| `tap`/`click` (touch vs. desktop) | Requires knowing whether the document targets a touch device, which Recheck cannot infer from text alone. |
+
+### DETECT-ONLY entries not shipped
+
+Real, confirmed guide content with no fixed replacement, excluded here
+because a generic "avoid this" pattern for these specific terms was judged
+lower-value than the DETECT-ONLY rules that did ship (`no-via`, `no-hover`,
+`utilize`, `performant`, `chapter-terminology`, `blind-figurative`):
+
+| Candidate | Why excluded |
+|---|---|
+| `persist` (transitive verb) | "Don't use as a transitive verb... best to avoid using as a verb at all" — no fixed replacement token, needs a rewrite. |
+| `comply`/`compliant` | Advisory caution only ("a claim that a product is compliant... is a strong statement"), no replacement at all — not a lint-shaped rule. |
+| `CLI` (bare, generic) | Needs "the specific CLI", which varies by context; no fixed swap target. |
+| `gray-box`/`graybox` testing | "describe exactly what it's doing" — no fixed term, and `translucent-box testing` is only an example, not a mandate. |
+| `tribal knowledge`/`wisdom` | "use a less figurative term" — no fixed replacement. |
+| `anti-pattern`, `canary`/`canarying` (verb), `best effort`, `out of the box` (figurative), `reservation, off the`, `voila` | Each is "avoid"/"don't use" with no fixed replacement token; the underlying content is confirmed but not independently high-value enough to ship as a detection-only pattern rule in this pass. |
+| `physically challenged`, `special`, `differently abled`, `handi-capable` | DETECT-ONLY per `inclusive-documentation`, no replacement offered. |
+| `right-hand side` (directional language, generally) | DETECT-ONLY, no fixed replacement — see also the TOO-RISKY note on `above`/`below` above; directional language generally was judged too broad to ship as its own rule (see "Author's judgment calls"). |
+
+### NOISY (verifier-confirmed content, judged too broad to enforce)
+
+**Accounting note, so the count below is checkable.** The task-9 corrections
+brief referred to "the 32 NOISY candidates from the research"
+(`research-google-style.md`'s own §4 tally). That figure is a count over a
+*different* population, using a *different* taxonomy, than this file's own
+four-row table below. The research draft numbered 146 candidate "rule
+ideas" and gave each one its own provisional, pre-verification classification
+(CLEAN/NOISY/NOT-ENFORCEABLE); 32 of those 146 rows carried NOISY as their
+first-stated verdict. This preset was not built by re-litigating those 146
+rows one-by-one — the five verification passes instead checked candidate
+content directly against the live guide's word-list and topic pages
+(~380 distinct entries, a finer unit than the 146 rows: e.g. one draft row
+can bundle a dozen word-list headwords). Cross-referencing the specific
+32 research-draft NOISY rows against what this file actually contains:
+
+- **6 ended up shipped anyway**, because a narrower scope, `fix: false`,
+  or a detection-only rule shape resolved the original noise concern
+  instead of requiring exclusion: `heading-sentence-case` (row 1, `fix:
+  false` + the `exceptions` mechanism), `no-code-in-heading` (row 9,
+  scoped to headings), `second-person` (row 13, detection-only),
+  `use-contractions` (row 17, `fix: false`), the timeless-documentation
+  words (row 29, 5 of ~15 shipped; see "Author's judgment calls" below for
+  the rest), and `copy-and-paste` (row 91, the copy/paste half only — the
+  same row's keyboard-shortcut suggestion was not shipped or excluded
+  anywhere, see below).
+- **5 correspond exactly to this table's own four rows** below (row 6 →
+  `no-gerund-headings`; row 52 → `no-slashes-general`; row 95 → table
+  sentence-case; rows 100 and 102 → the shared `no-quotes-around-code` /
+  `no-angle-brackets-around-code` row).
+- **2 correspond to NOT-ENFORCEABLE entries** below (row 39 → `oxford-comma`;
+  row 123 → gendered pronouns used generically).
+- **1 splits across two other tables**: row 145 ("console"/"CLI"/"UI"/
+  "Cloud"/"mobile" used bare) has its "CLI" part in DETECT-ONLY above and
+  its "Cloud" part in TOO-RISKY above; the "UI"/"mobile" parts of that same
+  row aren't reflected anywhere.
+- **The remaining 18 do not reappear anywhere in this file** — shipped,
+  excluded, or otherwise: nonbreaking space before a unit (row 69),
+  thousands-separator commas (row 61), avoiding seasons (row 65), SVG over
+  PNG (row 113), not forcing line breaks (row 115), inline HTML (row 116),
+  comma before "which"/"because" (rows 48-49), a colon instead of a dash to
+  introduce a list item (row 44), ambiguous-conjunction swaps like
+  since/while/once (row 131), modal-verb guidance (row 132), noun/verb form
+  pairs like setup/set up (row 133), spell-checking prose generally (row
+  142), foo/bar/baz placeholders (row 144), present tense / avoiding
+  will/would (row 16), the screen-reader punctuation caution (row 22), and
+  10x-style symbol substitutions (row 37). Their absence here means they
+  were never carried into the five verifiers' own ~380-entry check —
+  **not** that a verifier confirmed them and this preset silently dropped
+  them. Anyone who wants a disposition for one of those 18 specific ideas
+  should treat it as unverified against the live guide, not as any of this
+  file's excluded categories, and it is not shipped.
+
+The four rows below are this preset's own, narrower "NOISY" bucket: guide
+content a verifier independently confirmed as real, but judged too broad to
+enforce as a rule. That is a stricter, verifier-anchored sense of "NOISY"
+than the research draft's own pre-verification tally, which is why the
+counts don't and shouldn't match.
+
+| Candidate | Why excluded |
+|---|---|
+| `no-gerund-headings` (heading starting with an -ing word, except Billing/Pricing) | Confirmed principle, but the enforcement shape (first word ends in "-ing") would misfire heavily on common tech-noun headings used as topics, not verb-form imperatives (Networking, Logging, Caching, Monitoring, Testing) — Google names only two exceptions, which is itself evidence the underlying judgment needs more context than a heading's first word. |
+| `no-quotes-around-code`, `no-angle-brackets-around-code` | Confirmed, but the draft itself already flagged these as NOISY; not re-litigated. |
+| table sentence-case (`Use sentence case for all the elements in a table`) | Confirmed, but table cells often legitimately contain short labels, proper nouns, or numeric/code values that don't fit sentence-case cleanly (verifier B row 95, marked NOISY). |
+| `no-slashes-general` (`Avoid using slashes, except in code`) | Too broad — slashes appear constantly in dates, paths, fractions, and URLs; the draft itself marked this NOISY/dropped. |
+
+### NOT-ENFORCEABLE (real guide content, requires human judgment)
+
+| Candidate | Why excluded |
+|---|---|
+| `optional-prefix` (use "Optional:" prefix in headings) | Requires knowing whether a section is genuinely optional — not something the linter can determine. |
+| `complete-list-intro` (text before a colon must be a complete sentence) | Requires grammatical-completeness judgment beyond regex/AST primitives. |
+| `oxford-comma` (missing comma before the final "and"/"or" in a list) | Confirmed content, but reliably detecting a MISSING Oxford comma without a high false-positive rate on ordinary two-clause sentences needs real list/clause parsing, not regex. |
+| `abbrev-as-verb` (don't use acronyms as verbs, e.g. "ping the server") | Requires knowing a word's grammatical role (verb vs. noun), which the engine's primitives can't determine. |
+| `abbrev-first-use` (spell out an abbreviation on first mention) | Requires positional "is this the first mention" tracking across an arbitrary, unbounded set of abbreviations. |
+| `no-duplicate-link-destinations` (avoid linking one destination from different texts) | The guide states explicit exceptions (linking to a different section, a long page, multiple entry points) that the engine's existing token rule — which fires on same-destination/different-anchor-text — can't evaluate against. (The rule remains available generically via `recheck/markdown` for projects that want it unconditionally; it just isn't part of this style-fidelity preset.) |
+| external link icon (`Don't use an external link icon`) | About a rendered visual icon/CSS class, not markdown text — nothing to match. |
+| `ui-element-ellipsis` (drop the "..." from a UI element name reference) | Requires reliably identifying "this text is quoting a UI element name", which risks both over- and under-firing with a regex. |
+| `no-directional-language` (above/below/right-hand side as spatial UI references) | The guide's objection is to visual/spatial positioning language specifically; a text pattern can't distinguish that from the equally common non-directional uses of the same words (see the TOO-RISKY note above). |
+| `button-for-link` ("a link isn't the same as a button") | Requires knowing whether a referenced UI element is actually a button or a link, which text alone doesn't establish. |
+| gendered pronouns used generically (bare he/him/his/she/her) | Requires knowing whether a pronoun refers to a specific named person or is used generically — text alone can't distinguish. |
+| "introduce a table in the text preceding it"; the large "what belongs in code font" table; "don't pre-announce anything... unless approved by legal counsel"; "don't use metaphors"; passive voice ("make clear who's performing the action") | All require holistic judgment about content/structure/legal status that the engine's regex/AST primitives cannot evaluate. |
+
+### Dropped in Fix wave A (confirmed content, but not safely fixable/matchable)
+
+An independent review (task 9, fix wave A, 2026-07-29) reproduced four
+demonstrable prose-corruption defects and diagnosed a structural coverage
+gap that let 85 of the preset's 86 swap pairs ship with zero fixture
+coverage. These candidates were CONFIRMED guide content (they were already
+shipping) but are dropped here rather than fixed, because no safe
+detection/fix shape was achievable with the assertions available — see
+`task-9-report.md`'s "Fix wave A" section for the full defect list, every
+fix applied, and the re-run acceptance evidence.
+
+| Candidate | Why excluded |
+|---|---|
+| `'in line': 'inline'` (space-separated form, `google/compound-forms`) | Google's own quote ("One word as an adjective, inline, not in line or in-line") only objects to the ADJECTIVAL use, but bare "in line" is at least as commonly the correct idiom "in line with" or the plain verb phrase "wait in line"/"stand in line", neither of which the guide says anything about. Reproduced: `"This change is in line with the platform roadmap."` was being rewritten to `"...is inline with..."`. There's no reliable regex-only way to tell the wrong adjectival use apart from the idiom, so the key is dropped rather than shipped fixable or even detection-only. `'in-line': 'inline'` (the hyphenated form) is KEPT — it doesn't collide with the "in line with"/"wait in line" idioms, which are never written hyphenated. |
+| `'API Console': 'Google Cloud console'` (`google/product-names`) | Verifier B row 125 flagged that it doesn't map cleanly — the guide's own text offers "Google APIs Explorer **or** the Google Cloud console" by context, and which one is meant depends on what the original "API Console" reference meant. Not in verifier C's confirmed-clean list either. |
+| `Microservices: 'microservices'` (`google/acronym-forms`) | Fires on legitimate sentence-initial capitalization ("Microservices deployed on the platform can scale independently...") just as readily as on the actual violation (mid-sentence "the Microservices approach"), and `applyMatchCase` re-capitalizes the replacement to match the (all-caps-adjacent) matched casing, so the "fix" is a permanent no-op — an unsuppressible warning either way. There's no reliable way to detect "sentence-initial" from a `swap` pair alone. The lowercase/hyphenated form `'micro-services': 'microservices'` is KEPT (case-sensitive, so it never matches a legitimately-capitalized sentence start). |
+
+## Deliberation and development history
+
+Everything from here on records how the shipped and excluded lists above
+were reached: sourcing/hashing methodology, the fabrications an earlier
+draft introduced, the successive fix waves that corrected or excluded
+specific pairs, and the axes a pair must clear before auto-fix is safe.
+**A reader who only wants provenance can stop reading above this point** —
+everything below is for someone auditing or extending this preset.
+
+### `sources.json` normalization
+
+`sources.json` records one hash per source page as drift detection for a
+future re-check: re-fetch a page later, and a changed digest means the
+guide's content changed. The version of this file that shipped before
+this pass hashed the raw HTML response directly. That does not work as
+drift detection: every `developers.google.com/style/*` page embeds a
+per-request `<script type="application/json" analytics>` blob whose JSON
+keys serialize in non-deterministic order, a CSP `nonce` attribute
+regenerated on every request, and an inline feature-flag/experiment
+bootstrap array whose element order also varies per request. None of that
+reflects the guide's actual content, but it's enough entropy that two
+consecutive fetches of the identical page produce different raw-HTML
+digests every time — confirmed empirically: of the 30 pages listed in
+`sources.json`, fetched twice each a few seconds apart, 16 had a different
+raw byte length on the second fetch even though nothing about the guide's
+content changed. Hashing raw HTML gives a 100% false-positive drift rate,
+which is not drift detection at all — it's noise indistinguishable from
+signal.
+
+**Fix**: hash the extracted `<article class="devsite-article">...</article>`
+region instead of the full page. That region is exactly the guide's
+rendered content (headings, paragraphs, lists, tables, the `<dl>`
+definition lists the word-list page is built from) and excludes all of
+the non-deterministic chrome described above. Verified reproducible: two
+consecutive fetches of all 30 pages in `sources.json` (not just a sample)
+produced a byte-identical extracted region, and therefore an identical
+sha256, for every single page — including the 16 whose raw HTML length
+differed between fetches. Reproduction recipe: fetch the page's HTML with
+**`curl`** (Fix wave C / Item 4: a plain Node `fetch()` was confirmed to
+return a materially different HTML variant of the same URL — not just the
+per-request noise above, but a different response shape from the same
+client-vs-client comparison — so a future re-check that fetches with a
+different HTTP client could see a mismatch and misread it as guide drift;
+`sources.json`'s `normalization.fetchMethod` now records this too), take
+the first `<article class="devsite-article">...</article>` match (a
+non-greedy, dot-matches-newline regex scan; every page has exactly one),
+and sha256 the UTF-8 bytes of that substring.
+
+Each `sources.json` entry also records `bytes` — the byte length of the
+extracted region, not the raw page — so a future re-check that produces a
+matching `bytes` but a different `sha256` (or vice versa) is a signal to
+check whether the *extraction* shape changed (e.g. Google renamed the
+wrapper class) before treating the result as real guide drift.
+
+This normalization has a real, narrow blind spot: a change confined
+entirely to the `<article>` wrapper's own attributes, or to guide content
+that Google renders outside that element, would not be detected. No such
+case was observed across any of the 30 pages while producing this file.
+
+### Fix wave A corrections (2026-07-29)
+
+Beyond the drops above, several shipped pairs were corrected in place —
+same rule content, safer matching or `fix: false` — rather than dropped,
+because the underlying guide entry is real and worth keeping. Full
+before/after detail, reproduction commands, and gate output live in
+`task-9-report.md`; this is the pointer from provenance to what changed
+and why, so a rule split doesn't read as an unexplained addition.
+
+- **`google/no-latinisms` split in two.** The period-terminated keys
+  (`i.e.`, `e.g.`, `vs.`) keep `wordBoundary: false` (a trailing `\b`
+  right after a period-then-space never matches — both are non-word
+  characters) but now carry a LEADING `\b` baked into the regex source via
+  `keysAreRegex`, so `vs.` can no longer match inside unrelated words like
+  "revs.". `aka` and `vice versa` (no trailing period, so they never
+  needed the exemption) moved to a new rule, `google/no-latinisms-plain`,
+  with full `wordBoundary: true` anchoring — this is what stops `aka` from
+  matching inside "Akamai"/"Osaka".
+- **`google/no-slash-abbrev`** (`c/o`, `w/`) similarly gained a
+  leading-only `\b` via `keysAreRegex`, so `w/` no longer matches inside
+  "www/static", "show/hide", "new/old", and `c/o` no longer matches inside
+  "src/output".
+- **`google/acronym-forms`**: `'OAuth 2'` gained a `(?!\.0)` negative
+  lookahead so it can no longer match inside the already-correct "OAuth
+  2.0" (which was compounding into "OAuth 2.0.0.0.0.0.0" under
+  `runRulesUntilStable`). `SHA1` moved to its own rule
+  (`google/sha1-form`, `fix: false`) with a negative lookbehind excluding
+  a hyphen-preceded match, matching the guide's own documented
+  hyphenated-compound exception (e.g. "HMAC-SHA1"). `UNICODE` and `IPSEC`
+  moved to `google/acronym-caps-detect-only` (`fix: false`): both are
+  ALL-CAPS matches whose correct replacement (`Unicode`/`IPsec`), when
+  `applyMatchCase` re-upper-cases it to match the all-caps input, round-trips
+  back to the ORIGINAL wrong spelling byte-for-byte — a permanent,
+  silent no-op fix that `--fix` nonetheless reported as "fixed".
+- **`google/product-names`**: `'Cloud console'` gained a `(?<!Google )`
+  negative lookbehind — it's a literal substring of its own replacement
+  ("Google Cloud console"), so it was compounding a "Google " prefix every
+  pass, the same failure shape as the OAuth 2 bug above (and it fed the
+  same bug via `'Developers Console'`'s own correct output). `GCP` moved
+  to its own rule (`google/gcp-name`, `fix: false`): `applyMatchCase`
+  upper-cases an all-caps match's ENTIRE multi-word replacement, so
+  "GCP" -> "Google Cloud" was being written as "GOOGLE CLOUD".
+- **`google/colo-form`**: split out of `google/compound-forms`, `fix:
+  false`. `colo` is a noun ("a colocation facility"); `colocate` is a
+  verb — the unconditional swap produced "The colocate hosts the racks."
+- **`google/no-please-note`**: `fix: false` added. Deleting the phrase
+  leaves a capitalization/fragment mess behind ("Please note that the
+  endpoint is deprecated." -> "that the endpoint is deprecated.").
+- **`google/second-person`**: `ignoreCase: true` replaced with an explicit
+  `(?:We|we|Our|our|Us|us)` alternation, so it no longer matches the
+  all-caps abbreviation "US" that `google/us-abbreviation` fixes toward —
+  previously the two rules fought each other on every occurrence of "US".
+
+All of the above were found and fixed by the per-PAIR coverage gate added
+in the same pass (`src/config/__tests__/preset-google.test.ts`'s "per-pair
+coverage" test) plus targeted idempotency/self-match static analysis — see
+`task-9-report.md` for the methodology and the complete list.
+
+### Fix wave C corrections (2026-07-29)
+
+An independent re-review found this same `applyMatchCase` bug — an
+ALL-CAPS match forcing a multi-word replacement to shout — recurring a
+THIRD time (`AKA` -> `ALSO KNOWN AS`, `VICE VERSA` -> `THE OTHER WAY
+AROUND`, `C/O` -> `CARE OF`, all in rules that shipped with no `fix: false`
+workaround at all), after `GCP` and `UNICODE`/`IPSEC` above had each been
+patched around it per-rule. This wave fixed it at the source instead
+(`src/core/case-preserve.ts`'s `applyMatchCase`): an ALL-CAPS match no
+longer forces a MULTI-WORD replacement to upper-case; single-word
+replacements are unchanged (`WHITELIST` -> `ALLOWLIST` still shouts — that
+remains correct). Two statements above are now stale as a result:
+
+- **`GCP` is fixable again.** `google/gcp-name`'s `fix: false` (added
+  above specifically because `applyMatchCase` was shouting "GOOGLE CLOUD")
+  is REMOVED — the engine fix produces the correctly-cased `"Google
+  Cloud"` now, and the result is idempotent. `UNICODE`/`IPSEC`
+  (`google/acronym-caps-detect-only`) were re-checked against the same
+  engine fix and correctly stay `fix: false`: their replacements
+  (`Unicode`, `IPsec`) are each a single word, so the multi-word condition
+  never applies and the round-trip no-op is unchanged — a genuinely
+  different defect shape (same-word casing round-trip, not a
+  multi-word-phrase shout), not something this engine fix was meant to
+  reach.
+- **`google/product-names`'s `'Cloud console'` lookbehind widened.** The
+  `(?<!Google )` guard above only blocked the exact casing `Google `
+  (single space); `(?<![Gg][Oo][Oo][Gg][Ll][Ee]\s+)` now blocks any casing
+  of "google" followed by any run of whitespace, closing the same
+  self-compounding duplication (`"google Cloud console"` ->
+  `"google Google Cloud console"`) the original fix was meant to
+  eliminate but didn't fully.
+- **`google/no-slash-abbrev` gained a trailing guard.** The leading-only
+  `\b` above fixed matches inside a preceding word (`"src/output"`,
+  `"www/static"`) but left the TRAILING side open: `w/` still matched
+  inside the common, ordinary abbreviation `w/o` ("without"), and `c/o`
+  inside a word immediately following it (`"c/oscillator"`). Both keys now
+  carry a trailing `(?![A-Za-z])` negative lookahead blocking a following
+  ASCII letter, so `w/o` and `c/oscillator` are left alone entirely (the
+  guide's own `slashes` page doesn't list "w/o" as a separate entry),
+  while `"w/ headers"` and `"c/o the compliance department"` still match.
+
+See `task-9-report.md`'s "Fix wave C" section for the full list of every
+rule whose `--fix` output changed as a result of the engine change (9
+rules / 38 pairs across this preset alone — larger than the three rules
+named above, since the fix is in the shared helper every `swap` rule
+uses), the semantics chosen and why, and the complete acceptance evidence.
+
+### CONFIRMED vs. safe-to-fix — a distinction future audits of this preset must also apply
+
+Recorded here in mirror of `presets/microsoft/PROVENANCE.md`'s "Fix wave C /
+Step 5" (task 10), per that task's brief: a verifier `CONFIRMED` verdict
+establishes only that the live guide page discusses a term. It does **not**
+establish that a blind textual (`swap`) substitution of that term is safe
+to auto-apply — that is a separate question, and every verification pass
+in task 10 conflated the two. `recheck/microsoft`'s task-10 fix wave C
+found two pairs (`as well as` -> `and`, `or greater`/`or higher`/`or lower`
+-> `or later`/`or earlier`) that were marked plain `CONFIRMED`, with the
+correct quote attached, and still shipped as corrupting unconditional
+swaps: the quotes themselves showed a caution against treating two terms as
+interchangeable ("don't use X as a synonym for Y") or a context scoped
+narrower than what shipped ("when identifying multiple versions..."), not
+an unconditional "use Y instead of X" instruction.
+
+This note is a forward-looking record, not a claim that this file's own
+pairs were re-audited against that standard — that re-audit is out of
+scope for task 10 (which targets `recheck/microsoft`) and has not been
+done here. Whoever next adds a candidate pair to this preset, or authors
+either of the two future presets task 10 anticipates, should apply the
+same per-pair check `recheck/microsoft`'s Step 1 table used before marking
+a `CONFIRMED` pair fixable: is the live quote a direct "Use Y, not X"
+instruction, or does it instead read as a synonym-conflation caution, a
+verb/context-scoped rule, a multi-target rule, or a "don't use X" with no
+replacement stated? Only the first shape is safe to ship as an
+unconditional `swap`; the rest need a position anchor, a narrower scope, or
+detection-only, in that preference order. A `CONFIRMED` verdict is
+necessary evidence that a rule belongs in the preset at all — it is not
+sufficient evidence that the rule may safely auto-fix.
+
+**This note's own re-audit did eventually happen** — see "Fix-posture
+change" below, which re-audits every fixable pair in THIS preset against a
+second, orthogonal axis the note above doesn't cover, and replaces the
+"CONFIRMED vs. safe-to-fix" framing with a stricter, mechanical posture.
+
+### Fix-posture change (2026-07-30)
+
+> **RETIRED 2026-07-30 — see "Detection-only" below.** This section's
+> criterion (same-word normalization) no longer determines which pairs are
+> fixable in this preset: none are. Kept as historical record only.
+
+Base commit `eb4f8b11dac`, branch `aa/recheck-style-guides`. Brief:
+`.superpowers/sdd/preset-fix-posture-brief.md`. Report:
+`.superpowers/sdd/preset-fix-posture-report.md`. Applied to both flagship
+presets in the same pass — see `presets/microsoft/PROVENANCE.md`'s own
+"Fix-posture change" section for the full two-axis rationale and the six
+named corruption strings that motivated it (three fix waves on
+`recheck/microsoft`, each fixing the pairs a probe found, each followed by
+another probe finding more — 2, then 2, then 6). This preset shipped nine
+of the same defect class across its own three fix waves (`GCP` shouting,
+`OAuth 2` self-compounding, `w/o`/`c/oscillator` corruption, ...) — the
+CONFIRMED-vs-safe-to-fix note above addresses the GUIDANCE-SHAPE axis
+(direct instruction vs. caution/scoped/multi-target); it does not address
+a second, orthogonal axis: whether the avoid-term ALSO has a legitimate,
+unrelated sense a blind substitution corrupts regardless of how clearly
+the guide states its rule.
+
+#### The posture
+
+Auto-fix is retained only where a replacement cannot be wrong: the same
+word, normalized (spelling, hyphenation, casing, or a non-standard written
+form of the identical word). A pair that substitutes a genuinely different
+word or phrase — even a synonym that looks safe on the page — moves to
+detection-only. Concrete examples found THIS wave, previously shipped
+fixable under a plain `CONFIRMED`/reasonable-looking verdict:
+
+- **`agnostic` → `platform-independent`** (`google/plain-language-swaps`):
+  "agnostic" very commonly means doubting or noncommittal about religious
+  or philosophical claims ("he's agnostic about the existence of an
+  afterlife") — a blind fix corrupts that sentence into "...platform-
+  independent about the existence of an afterlife." The word-list entry
+  is real and the replacement is reasonable for the INTENDED sense; it is
+  simply also a different word with an unrelated common sense, the same
+  shape as `recheck/microsoft`'s `DMZ`.
+- **`GCP` → `Google Cloud`** (`google/gcp-name`): the same shape as `DMZ` →
+  `perimeter network` — an acronym expanded into a DIFFERENT phrase than
+  its own literal expansion (`GCP` stands for "Google Cloud Platform", not
+  "Google Cloud"), not a respelling, and `GCP` has unrelated expansions in
+  other domains ("Good Clinical Practice", "Grade Control Point"). Note
+  this reverses Fix wave C's own "GCP is fixable again" call above: that
+  wave correctly fixed a real ENGINE bug (the case-preservation shout),
+  but engine-correctness and word-choice-safety are different questions,
+  and only the first was checked at the time.
+- **`IO` → `I/O`** (`google/acronym-forms`, moved to
+  `google/acronym-caps-detect-only`): bare, case-sensitive "IO" is a real
+  product/library name in common developer use ("Socket.IO" — the period
+  before "IO" is a non-word character, so `\bIO\b` matches inside it) —
+  "Socket.IO connects clients" would corrupt to "Socket.I/O connects
+  clients."
+
+By contrast, `google/compound-forms`'s ~80 remaining pairs (`data store` →
+`datastore`, `e-mail` → `email`, ...) are genuine spacing/hyphenation
+normalizations of the identical two words and stay fixable — this is the
+family the fix-posture brief itself predicted would "largely survive."
+Four pairs did NOT survive despite living in that same rule: `data
+cleansing` → `data cleaning` (different word, "cleansing" vs "cleaning"),
+`transcompile` → `transpile` (two competing compiler-jargon terms, not a
+spelling variant), `autoupdate` → `automatically update` (expands "auto"
+into a different word), and `pre-emptive` → `preemptible` (a different
+adjective — "pre-emptive" describes acting in advance; "preemptible"
+describes being subject to preemption — not a hyphenation of one word).
+Moved to a new sibling rule, `google/compound-forms-word-choice`
+(`fix: false`), since `fix` is a whole-rule flag and this bundle mixed
+both classes.
+
+#### Result
+
+**Fixable `swap`/`consistency` pairs: 164 → 114**, across 41 rules with a
+`swap`/`consistency` assertion (up from 38 — `google/vs-versus`,
+`google/aka-form`, and `google/compound-forms-word-choice` are new,
+splitting bundles that mixed same-word and different-word pairs).
+`vs.` → `versus` and `aka` → `also known as` both stay fixable: unlike
+`i.e.`/`e.g.` (Latin abbreviations translated into an unrelated English
+phrase) and `vice versa` (a distinct Latin phrase with no letter-derived
+relationship to its replacement), `vs.` and `aka` are literal truncations
+of the identical word/phrase they abbreviate (`aka` is literally the
+initials of "Also Known As"). **Correction (wave 2, see below): this call
+on `aka` was wrong.** Expanding an abbreviation INTO a phrase is a
+substitution, the same shape as `e.g.`/`i.e.` two sentences above, not a
+respelling — "the letters spell out the words" does not make it a
+same-word normalization. `aka` moved to `fix: false` in wave 2. `cURL` →
+`curl` (a pure casing correction)
+moved from `google/product-names` (now fully detection-only) to
+`google/brand-capitalization`, which already ships the same class of
+case-only brand-name fix.
+
+Severity is unaffected: this preset's existing policy already puts every
+word-choice/phrasing rule at `warn` regardless of fixability (see "Shipped
+rules" above), so newly-detection-only rules needed no severity change.
+
+#### Gates
+
+`pnpm build`, `pnpm test` (105 files, 1467 passed / 5 skipped — shared
+suite with `recheck/microsoft`), `pnpm parity --corpus monorepo-docs`
+(unchanged: 27425 = 27425, 0 unexplained — this preset has no bearing on
+the markdownlint-parity corpus), `npx nx run recheck:lint
+--max-warnings=0` (clean). Every one of the 114 pairs that remain fixable
+in this preset (240 combined with `recheck/microsoft`'s 126) was verified
+programmatically: real change, idempotent (second `--fix` pass is a
+no-op), and the original violation regex no longer matches the fixed
+text — not sampled. Full command output and the exhaustive fixable-rule
+table are in `.superpowers/sdd/preset-fix-posture-report.md`.
+
+### Fix-posture change, wave 2 — the proper-noun axis (2026-07-30)
+
+> **RETIRED 2026-07-30 — see "Detection-only" below.** This section's
+> third axis (proper-noun collision) no longer determines which pairs are
+> fixable in this preset: none are. Kept as historical record only.
+
+Base commit `25a62c3d1f1`, branch `aa/recheck-style-guides`. Brief:
+`.superpowers/sdd/preset-posture-fix2-brief.md`. Report:
+`.superpowers/sdd/preset-fix-posture-report.md`'s "Wave 2" section. See
+`presets/microsoft/PROVENANCE.md`'s own "Fix-posture change, wave 2"
+section for the full rationale shared by both presets; this section
+covers this preset's specific findings.
+
+#### The third axis
+
+A pair keeps `fix: true` only if it is the same word normalized (axis 1,
+guidance-shape), has no unrelated legitimate sense (axis 2, homograph —
+wave 1), **and, new this wave, cannot occur as part of a real
+organization, product, brand, or place name.** Four demonstrated
+corruptions share this cause: `markdown` → `Markdown` (a retail markdown
+sentence gets capitalized into the markup-language name), `FinTech` →
+`fintech` (breaks "FinTech Group AG", a real company), `U.S.` → `US`
+(breaks "U.S. Bank", a real bank), `USA` → `US` (breaks "USA Gymnastics",
+a real governing body — this specific pair lives in
+`microsoft/usa-abbreviation`, not this preset, since `google/us-
+abbreviation` never shipped a bare `USA` key). Acronyms and single
+capitalizable words are the highest-risk shapes: a rule that ignores
+case or normalizes punctuation matches a proper noun's own official
+spelling exactly as readily as ordinary prose.
+
+#### Sweep and results
+
+Every fixable `swap` pair in this preset (105 pairs after this wave,
+sweeping the full ~114-pair fixable set left by wave 1) was checked
+against question 3.
+
+| Rule | Pair(s) flipped | Real proper noun / other reason |
+|---|---|---|
+| `google/aka-form` | `aka` → `also known as` | Not proper-noun — **misclassified in wave 1** (see the correction note above): expanding an abbreviation into a phrase is a substitution, the same shape as `e.g.`/`i.e.`, not a same-word normalization. Also happens to double as a real proper noun ("AKA" is the common abbreviation for the sorority Alpha Kappa Alpha — "She was initiated into AKA her freshman year" is a genuine, independent proper-noun collision on top of the misclassification). |
+| `google/us-abbreviation` | `U.S.A.`/`U.S.` → `US` (both) | "U.S. Bank" (top-10 US bank), "U.S. Steel", "U.S.A. Track and Field" (national governing body) — the named brief case. |
+| `google/acronym-forms` | `FinTech` → `fintech` | "FinTech Group AG" — a real, publicly-traded German company whose name keeps the mixed-case "FinTech" spelling. |
+| `google/acronym-forms` | `I-O` → `I/O` | "I-O DATA DEVICE, INC." — a real, major Japanese PC-peripherals manufacturer ("Japan's undisputed market leader" in that industry) whose brand is written exactly "I-O" (hyphenated, capital letters). `fin-tech`/`adtech`/`ad-tech` (all-lowercase keys, no case change involved) do NOT carry this risk: this rule is case-sensitive, so an all-lowercase key can never match a capitalized brand's own casing in the first place. |
+| `google/brand-capitalization` | `markdown` → `Markdown` | The named brief case — both a retail/finance homograph (axis 2) AND, in the reverse direction, a proper-noun risk: capitalizing every lowercase occurrence assumes it always means the Markdown language. |
+| `google/brand-capitalization` | `material design` → `Material Design` | Same reverse-direction risk as `markdown`: "the material design of the building incorporates local stone" is a plain, unrelated architectural phrase this pair would wrongly capitalize into Google's design-language name. |
+| `google/brand-capitalization` | `search console` → `Search Console` | Same shape, weaker but still real: a generic lowercase phrase for an admin/tuning panel, not exclusively Google's product name. |
+| `google/compound-forms` | `datasource` → `data source` | Collides with `javax.sql.DataSource`/Spring's `DataSource` — a real, load-bearing Java/Spring class and config-property name ("Configure the DataSource bean in the Spring context"), exactly the kind of technical content this preset's own audience writes about. |
+
+**9 pairs flipped**, each moved to a new detection-only sibling rather
+than anchored (`google/acronym-forms-proper-noun`, `google/brand-
+capitalization-proper-noun`, `google/compound-forms-proper-noun`;
+`google/aka-form` and `google/us-abbreviation` flip whole-rule since every
+pair in each was reclassified). Severity is unaffected — this preset's
+existing policy already puts every word-choice/phrasing rule at `warn`
+regardless of fixability, so nothing needed adjusting, matching wave 1's
+own note.
+
+**Fixable pairs: 114 → 105.**
+
+#### Why `fix: false`, not another anchor
+
+Same reasoning as `recheck/microsoft`'s wave 2 section: three separate
+waves have each shown an anchor leaking exactly one near-miss beyond
+wherever it was tested. A casing/abbreviation fix is low-value enough
+that detection alone is a fine outcome, so every pair this wave found
+moved straight to `fix: false` rather than growing another exclusion
+list.
+
+#### Gates
+
+`pnpm build`, `pnpm test` (105 files, 1490 passed / 5 skipped — 23 new
+tests this wave, shared suite with `recheck/microsoft`), `pnpm parity
+--corpus monorepo-docs --profile default` (unchanged: 27425 = 27425, 0
+unexplained), `npx nx run recheck:lint --max-warnings=0` (clean). Every
+sentence in the brief's acceptance set 1 verified unchanged through
+`--fix` twice and still detected against the correct new rule name; every
+rule (including the 3 new ones here) still fires on
+`google-violations.md`. Full output in
+`.superpowers/sdd/preset-fix-posture-report.md`'s "Wave 2" section.
+
+### Detection-only (2026-07-30)
+
+Base commit `006c026a1f0`, branch `aa/recheck-style-guides`. Brief:
+`.superpowers/sdd/preset-detection-only-brief.md`. Report:
+`.superpowers/sdd/preset-detection-only-report.md`.
+
+**This section REPLACES the fixability criterion described in "CONFIRMED
+vs. safe-to-fix," "Fix-posture change," and "Fix-posture change, wave 2"
+above — it does not sit alongside them as a fourth, stricter axis.** Those
+three sections are kept below, unedited, as the historical record of the
+criteria that were tried and superseded; do not read any of them as current
+guidance. As of this section, **`recheck/google` ships zero fixable rules.
+Every rule in this file is `fix: false`, unconditionally** — set
+structurally, once, by a loop at the end of `buildGooglePreset()`
+(`src/config/presets/google.ts`), not by auditing pairs against a sharper
+rule. A dedicated test (`preset-google.test.ts`'s "is detection-only"
+describe block) reads the live preset object and fails if any rule is ever
+fixable again — the same derive-from-the-preset shape the per-pair coverage
+gate already uses, so this cannot regress silently the way three prior
+narrowing passes did.
+
+#### Why a fourth axis wasn't the answer
+
+Two prior fix-posture changes (above) each replaced "does the guide confirm
+this pair" with a sharper structural test — first "is it the same word
+normalized" (axis 1: guidance-shape, then axis 2: homograph), then "does it
+also collide with a real proper noun" (axis 3). Each pass shipped clean
+against its own criterion and each was then probed again. The fifth
+adversarial probe against this preset and `recheck/microsoft` together (the
+project's fifth in total, after three rounds already narrowed what counted
+as "safe") found **18 of 29 probed pairs (62%) still corrupting correct
+prose** — a RISING hit rate, not a falling one, and the failures spanned
+every category previously believed safe by axes 1-3, including two this
+preset's own criteria treated as clean:
+
+- **Spelling**, believed the safest category of all: Hemingway's real,
+  correctly spelled published title *A Moveable Feast* is corrected to "A
+  Movable Feast" by `microsoft/az-grammar-usage`'s `moveable` → `movable`
+  pair (a genuine same-word normalization by every axis above — axis 1
+  passes, axis 2 finds no unrelated sense, axis 3 finds no proper-noun
+  collision on the WORD "moveable" itself, and the collision is instead
+  with a specific, individually unforeseeable literary title).
+- **Hyphenation**, this preset's own `read only` → `read-only` pair
+  (`google/compound-forms`): "Please read only the introduction" — an
+  adverb ("only") modifying a verb ("read") plus its object — becomes
+  "Please read-only the introduction," a nonsense adjective use. Same-word
+  by every axis (it is the identical two words, just joined), yet wrong,
+  because the axes check the WORDS, not the GRAMMATICAL ROLE those words
+  are playing in the sentence being fixed.
+- **Meaning inverted outright**: `google/acronym-forms`'s `No SQL` → `NoSQL`
+  turns "No SQL is used here" (a true statement that no NoSQL database is
+  in use) into "NoSQL is used here" (a false statement that one is) — same
+  word-pair, same axis-1/2/3 clearance, opposite meaning.
+
+Full round-5 acceptance evidence (every sentence above, and more, run
+through `--fix` twice and confirmed byte-identical) lives in
+`src/config/__tests__/preset-detection-only-acceptance.test.ts`, plus the
+per-preset regression suites in `preset-google-fix-wave-c.test.ts` and
+`preset-microsoft.test.ts` (both rewritten by this change to assert
+"unchanged" where they used to assert a real rewrite).
+
+#### The conclusion this decision rests on
+
+A rule's *category* — spelling, hyphenation, casing, word-choice — does not
+predict fix-safety at this scale. A style guide states *intent* ("use X to
+mean Y"); a `swap`/`consistency`/`pattern` rule matches *tokens* (literal
+text, regardless of the grammatical role or referent that text has in a
+given sentence). That gap is not closable by inventing a fourth, fifth, or
+sixth axis: axis 1 (guidance-shape) closed the space of pairs where the
+guide's own wording was ambiguous; axis 2 (homograph) closed the space of
+words with an unrelated common sense; axis 3 (proper-noun) closed the space
+of words that double as real names. Each closure found the NEXT gap, not
+zero gap. The project decision is to stop narrowing and remove fixing
+capability from both style-guide presets entirely: users get every finding
+(detection is completely unaffected — every rule still runs `execute()` and
+reports) and apply the judgment a style guide has always required, same as
+before either preset existed and same as Vale (the tool these presets
+replace), which never shipped an auto-fixer and never had this class of
+bug.
+
+#### What did not change
+
+Detection. Every rule's `execute()` path, message, severity, and scope are
+untouched — only `fix()` is gated off (`core/runner.ts`'s
+`rule.fix !== false` check). The per-pair and per-rule coverage gates
+(`preset-google.test.ts`) still require every rule to fire on its own clean
+fixture, so a rule that neither fixes nor reports is still caught as dead
+weight, same as before this change.
+
+### Known limitations
+
+Two verifier-confirmed guide-sanctioned exceptions that the shipped rules
+do not implement. Both are documented here explicitly, not just in the
+Shipped rules table's own one-line notes, because a rule that is stricter
+than its source needs to say so where a reader is actually looking for
+"why did this fire on text the guide allows" — otherwise a user hitting
+either case reasonably concludes the preset misquotes Google.
+
+1. **`google/no-emphasis-as-heading` over-fires on one shape of the
+   guide's own "run-in heading" pattern (verifier A row 12).** The guide
+   permits bold for "run-in headings" — a bolded lead-in term followed by
+   its description, most commonly inside a description-list item (`Google's
+   own example: <li><b>Emu</b>: the best kind of bird</li>`) or a single
+   paragraph ("**Emu:** the best kind of bird."), and instructs authors to
+   "end the run-in heading with a period or a colon." The shipped rule
+   (ported from markdownlint's MD036) already tolerates the common cases:
+   it only examines top-level paragraphs (so a run-in heading inside an
+   actual list item is never even considered), a bold lead-in with more
+   text in the SAME paragraph is excluded (the paragraph has more than one
+   meaningful child), and a bold-only paragraph ending in the guide's own
+   required punctuation is excluded by the rule's pre-existing punctuation
+   check. Empirically verified still-flagged: a bold/italic-only paragraph
+   with NO ending punctuation, whose description follows in a SEPARATE,
+   later paragraph (e.g. `**Emu**\n\nThe best kind of bird.`) — reproduced
+   directly against the shipped rule. Not fixed in this pass: the rule is
+   a markdownlint port shared with `recheck/markdown` (out of scope for
+   this wave's provenance/documentation focus, and doing so risks the same
+   loosening-vs-noise tradeoff every other TOO-RISKY exclusion in this file
+   weighs).
+2. **`google/no-url-as-link-text` does not exempt legal/ToS documents
+   (verifier B row 71).** The guide's own text: "Exception: In some legal
+   documents (such as some Terms of Service documents), it's okay to use
+   URLs as link text." The shipped rule is a plain pattern match on link
+   text starting with `http(s)://`, scoped to `link`; it has no way to
+   know whether the document containing the link is a legal/ToS document,
+   so it will flag a bare URL used as link text there too, against the
+   guide's own stated exception. Not enforceable to fix with the engine's
+   current primitives (same class of gap as `firewalls` → `firewall
+   rules`'s "Compute Engine documentation only" scoping and the
+   NOT-ENFORCEABLE table's document-subject-matter entries above) — Recheck
+   has no signal for what kind of document a file is.
+3. **`google/gcp-name` is case-sensitive.** The pair (`GCP: 'Google
+   Cloud'`) carries no `ignoreCase`, so only the literal all-caps `GCP`
+   token is matched or fixed — `gcp` and `Gcp` are neither flagged nor
+   fixed by this rule. A reader could reasonably infer broader coverage
+   than exists: many OTHER rules in this same file (e.g. `google/
+   compound-forms`, `google/use-contractions`) do set `ignoreCase: true`,
+   so the absence here is easy to read as an oversight rather than a
+   choice. It is a choice, shared with the sibling `google/product-names`
+   (also no `ignoreCase`) and `google/brand-capitalization` (explicitly
+   documented as "deliberately case-sensitive keys, matching only the
+   wrongly-cased literal form" in its own comment) — all three treat
+   brand/product-name casing as exact-match by design. Left as-is rather
+   than widened here (carried over from a `recheck/microsoft` fix-wave
+   audit, flagged as a documentation gap, not a behavior bug): widening to
+   `ignoreCase: true` is a scope decision for whoever owns `google.ts`, not
+   a documentation fix.
+
+### Author's judgment calls
+
+Decisions this preset's author made that go beyond a verifier's literal
+verdict, recorded per the task's instruction to flag (not silently
+resolve) anything not settled by the inputs:
+
+1. **`google/first-line-h1` and `google/single-h1` share one Google
+   quote.** Verifier A's row 4 confirms both rule ids against the same
+   sentence ("only use a level-1 heading once on a page"); the research
+   draft, not a separate guide statement, is what split them into two rule
+   ids. Both ship (both are real markdownlint-ported mechanisms Google's
+   principle supports), but see point 2.
+2. **`single-h1` and `first-line-h1` cannot both fire from one document.**
+   Empirically verified (not assumed): the underlying token rules faithfully
+   port markdownlint's MD025/MD041, which check OPPOSITE preconditions of
+   "the document's first heading" — `single-h1` only reports a second h1
+   when nothing but comments/frontmatter precede the first one;
+   `first-line-h1` only fires when that first real content is NOT a correct
+   h1. `google-violations.md` (which starts with a level-2 heading to
+   trigger `first-line-h1`) structurally cannot also trigger `single-h1`, so
+   a second, tiny fixture (`google-violations-single-h1.md`) isolates it.
+   This is a genuine engine/upstream-semantics interaction, not a fixture
+   bug or a noisy rule.
+3. **Downgraded `list-length` from the generic "list mechanics = error"
+   class to `warn`.** The verifier marked the underlying quote
+   NOT-ENFORCEABLE (descriptive, not imperative); only the mechanism
+   (`list-length`'s `min: 2` default) is deterministic, so the softer
+   severity reflects the guide's own softer confidence.
+4. **Shipped only 5 of ~15 confirmed "timeless documentation" words.** All
+   ~15 are confirmed content, but 10 of them (`existing`, `future`,
+   `latest`, `new`, `newer`, `now`, `old`, `older`, `soon`, `eventually`,
+   `in the future`) are ordinary high-frequency English words with
+   extensive legitimate everyday use unrelated to documentation staleness.
+   Shipping them would make the preset unusably noisy on typical prose —
+   the same class of risk the verifiers flagged elsewhere as TOO-RISKY,
+   extended here on the same reasoning to entries the verifiers didn't
+   individually re-litigate for riskiness (their job was confirming
+   content, not judging blind-match safety for every term).
+5. **`google/no-numbered-headings` ships despite some residual risk.**
+   Narrowly scoped to `Step N`/`Part N` markers and a bare leading ordinal,
+   to keep the false-positive rate low; broader numeric heading patterns
+   (e.g. version numbers in a heading) are not matched.
+6. **`google/cons-and-pros` ships only the compound phrase.** Bare `pros`/
+   `cons` are excluded even though individually confirmed, because standing
+   alone they're closer to ambiguous (conference abbreviations, "con
+   artist") than the extremely common, unambiguous two-word phrase.
+7. **The Oxford-comma/no-and-or "except in tables" exception is not
+   separately scoped.** `google/no-and-or` runs over the `summary` scope,
+   which includes table cells, so it would also (correctly, per the general
+   rule, but against the guide's own table exception) flag "and/or" inside
+   a table. Judged not worth a bespoke scope array for one rule; a project
+   that hits this can override the rule's `scope`.
+
+### Engine/registry changes this preset required
+
+- **Schema**: `src/config/schema.ts`'s top-level `patternProperties` only
+  accepted `^recheck/[a-z0-9-_]+$` rule keys. Spec §2 ("Composition
+  safety") requires per-preset namespacing (`google/<rule>`,
+  `microsoft/<rule>`, ...) precisely so two flagship presets can be
+  composed without collisions — the pattern is widened to
+  `^[a-z][a-z0-9-]*/[a-z0-9-_]+$` to allow that (existing `recheck/*` keys
+  are unaffected; they're just the `recheck` namespace now).
+- **`length` moved from opt-in to preset-shipped.** `google/sentence-length`
+  is the first non-prose preset rule to ship a native scope-rule
+  assertion beyond what `recheck/prose` already ships. Per
+  cross-task-constraints.md §C / task-9-10-resolutions.md §5, this trips
+  the registry<->preset completeness guard in
+  `src/config/__tests__/presets.test.ts`: `length` is removed from the
+  documented-opt-in list (now `DOCUMENTED_OPT_IN_ASSERTIONS`, moved from
+  `prose.ts` to `presets/index.ts` since the policy is monorepo-wide, not
+  prose-specific) and the completeness test's "shipped" side is derived
+  dynamically from ALL presets rather than a single prose-named constant.
+  See that test file's own comments for the mechanics.
+- **`.npmignore` widened to include `presets/**/*`.** The package has no
+  `files` field in `package.json`; publishing is governed entirely by
+  `.npmignore`, which was a blanket `*` deny with only `dist/**/*` and
+  `package.json` allowed back in. A new top-level `presets/<name>/`
+  directory (this file, `sources.json`) would have shipped nowhere without
+  this change — verified with `npm pack --dry-run` before and after.