sheleg-design-skill 1.45.0 → 1.46.0

package/CHANGELOG.md

@@ -4,6 +4,1172 @@ All notable changes to this project are documented in this file. The format
 follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions
 follow [SemVer](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+## [1.46.0] - 2026-08-22
+
+### `babylove` — the thirty-first pack, and the other end of the same question
+
+Measured 2026-08-21 off <https://www.babylovegrowth.ai/en> and its
+`/dashboard/overview` in a logged-in session.
+
+**The pair is the argument.** `outrank` and `babylove` are the same product category in the
+same year — SEO and AI-visibility SaaS with a landing and a dashboard. `outrank` answers *how
+much system do you need* with **536 custom properties across two borrowed design systems**.
+This one answers it with **seven**: `--brand` and six steps of it, every grey taken from
+Tailwind v4's defaults, and nothing else declared at all. Neither is wrong, and the library now
+carries both ends priced.
+
+Three findings the arithmetic produced:
+
+* **The brand orange cannot carry a word in either direction** — `#FA5C12` is 3.18:1 on the
+  page and white on it is 3.18:1. The reference knows: its hero CTA is **near-black at
+  19.66:1**. The one place it breaks its own rule is the nav button, white on `#F25533` at
+  **3.43:1**, in the same viewport as the one it got right.
+* **The rail's section labels are 2.60:1** — Tailwind's gray-400 at 11px, and 2.49:1 on the
+  inner card. Corrected to gray-500 `#6A7282`, which is the framework's own next step rather
+  than a value from me.
+* **`--brand-dark` is 4.49:1** — one hundredth under AA. Kept because the reference declares
+  it, and named as not-the-token-for-text.
+
+Its signature is the **disconnected state as a first-class surface**: a card whose source is
+not connected renders at full size with an illustration, a sentence naming what connecting
+buys, and a button — so the dashboard's shape does not change when a source is added, and a new
+account reads as unfinished rather than as broken.
+
+### Two gate defects the thirty-first pack exposed
+
+**`NUMBER_WORDS` stopped at thirty, and a second copy of it stopped there too.**
+`COUNT_WITHOUT_ITS_NOUN` spelled the same vocabulary out a second time. At thirty-one the first
+turned every count check red — loudly, correctly — and the second **went quiet on exactly the
+sentence it exists for**. The silent one is the dangerous half, and it surfaced only because
+its own plant then reported the validator going red for an unrelated reason. The pattern is
+built from the dictionary now, longest-alternative-first so `twenty` cannot shadow
+`twenty-one`.
+
+**Four plants are anchored on counts that every new pack moves.** Adding `outrank` broke two;
+adding `babylove` broke all four. Each pinned the literal a count edit rewrites. All four
+derive their anchor now. The residue — 57 plants not read for the same shape — is B-131.
+
+### `outrank` — the thirtieth pack, and the first that is two surfaces
+
+Measured 2026-08-21 off <https://www.outrank.so> and its `/dashboard` in a logged-in
+session, with `getComputedStyle` in both themes.
+
+**Said before anything else: none of this is Outrank's invention.** The marketing site runs
+**Tailwind + daisyUI** — `--rounded-box`, `--btn-text-case` and the `--p`/`--pc` HSL triplets
+are daisyUI's own names — and the product runs **Semrush's Intergalactic**, 384
+`--intergalactic-*` properties on the dashboard. A pack presenting this as a bespoke language
+would be claiming authorship of two published systems.
+
+That is the reason to have it. **Every other pack in this library is one surface.** This one is
+the seam between a landing that must convert and a tool that must be worked in for hours, and
+the two halves *disagree* — on the ink, the radius and the type family — while agreeing
+completely on the brand. A team shipping both needs that agreement written down.
+
+What the measurement produced, and it is not decoration: **the pack corrects the reference
+three times, and each correction is the smallest the rule permits.**
+
+* The eyebrow `#64748B` is **4.37:1 on `--surface`** at the 13px it is set at — below AA on
+  the field it appears on most. `#5B6779` is the nearest value clearing 4.5 on both fields.
+* The warning `#FF642D` is **2.95:1**, under even the 3:1 non-text floor. Darkening alone was
+  not enough: at `#E5511B` it sits **7.2 OKLab from `--danger`** against a hard floor of 10.0,
+  so red and orange are one state to a reader who cannot separate them. `#C2740A` is an amber
+  15.2 away — and amber is the conventional warning encoding, so separation and convention
+  agree.
+* The brand violet **does not survive the dark theme** at 2.88:1 on the dark card. The twin
+  uses the reference's own `--violet-400` `#AB6CFE` at 4.76:1 — taken from its scale, not
+  invented.
+
+And one the reference gets right that most packs get wrong: `--accent` `#882AFF` carries text
+**in both directions** — 5.49:1 as a word on the page and 5.49:1 under white on the filled
+button. Most brand violets at that saturation clear neither, which is why the value is kept.
+
+The kit ships eight components on the `or-` prefix, including a `DataTable` whose header row is
+the eyebrow carrying its own count, and an `EmptyState` that keeps its border — on the
+reference's content calendar an empty day still renders its cell, because the month's shape is
+the information.
+
+**Disabled is a token pair here, not an opacity.** WCAG 1.4.3 exempts an inactive control, which
+is why most packs reach for `opacity: 0.38` — the palette gate computes that composite and
+refuses it. `--ink-eyebrow` on `--surface-2` is 4.40:1, so a disabled control still reads as the
+control it is.
+
+Registering the thirtieth pack moved 23 files' counted claims from twenty-nine to thirty and
+broke two plants whose anchors were those very numbers — caught by the self-test and
+re-anchored, which is the failure mode this repository already had on record.
+
+### Five defects the gate could not see, found by three blind readings (B-125…B-130)
+
+`task-pipeline` v1.74.0 closes a node with three independent readings at escalating visibility
+instead of one verdict. Its **first live dispatch** was aimed at this repository's own
+`f88c14b` — the strict reduced-motion gate — and it returned two `fail` verdicts and five
+`breaks` findings against work that had shipped green. What each tier caught is the argument for
+the mechanism, so it is recorded here rather than summarised:
+
+* **The gate matched durations by NAME, and the product tier proved it with a plant.** It removed
+  `--t-hero-art: 0s;` from `paperclip`'s reduced-motion branch in all three copies, left
+  `--t-hero-art: 1.1s` in `:root`, and watched the gate print `0 silent duration(s)` and exit 0.
+  Thirteen time-valued tokens across ten layers were outside the walk entirely — `paperclip`'s six
+  `--t-*`, seven `--stagger*`, `--marquee-cycle`, `--scan-period` — all answered today by the
+  authors' habit rather than by the check. **It matches by value now: 127 durations examined, was
+  112**, and that plant is kept so the name-keyed form cannot return.
+* **The ratchet priced the exception rather than the rule.** The seam tier instrumented `check()`
+  and collapsed `atrium`'s four kept durations — the remediation the requirement names *first* —
+  and measured the count **drop by 4 against a floor of 4026**, turning `npm test` red on the
+  stricter answer, where the only way through was lowering a ratchet whose own `_why` says a
+  falling count is how a deleted requirement hides. The walk now contributes **exactly one check
+  per duration** whatever its answer, so the floor is a property of the corpus.
+* **`docs/DOCMAP.md`'s propagation row still stated the obligation the commit replaced** — one
+  collapsed duration, and the superseded check as its proof. **Both blind tiers found this
+  independently**, which is the clearest signal the two levels are reading different things. An
+  author obeying the row failed the gate; the commit had edited the CHANGELOG, the board, the
+  floors, the gate and the CSS, and not the table that tells a contributor what to do.
+* **A published pair was restated and wrong.** *8 of the library's 11 re-declarations* resolved
+  against no reading of *re-declaration* the document defined. Measured: **112 at `0s`, 9 at the
+  0.01ms idiom, 6 above an instant.** Both figures are computed here now.
+* **B-120's residual scope did not resolve in the repository holding it** — it named
+  `test/*_test.js` and *the other three ratchet files*, and there are exactly three ratchet gates,
+  all covered, and no `*_test.js` in this tree.
+
+Four more findings are filed rather than fixed, because each is a decision rather than a repair:
+the re-declare-with-reason form is enforced by a gate and sanctioned by no document (**B-126**);
+`atrium`'s canvas-removal promise is unverifiable here and reads as covered (**B-127**); `prism`'s
+ambient point cloud has no stated still (**B-128**); the reason window crosses reduce-block
+boundaries (**B-129**); and four competing definitions of *a duration token* now live in one file
+(**B-130**). **B-125** is the one about me: every floor set on 2026-08-20 sits below its true
+count, because the sequence was measure → keep editing → restate.
+
+### The pairing check the prose ratios needed had been unreachable since it was written (B-053)
+
+Forty-six stated ratios sat in prose that named no partner, and the board carried them as a
+residue to be closed one sentence at a time. **The check they needed already existed.** Twenty-nine
+lines inside `validate_stated_ratios` — a full pairing pass that narrows to the token named on the
+line and refuses a number no pair produces — sat behind a `continue`, unreachable from the day it
+was written, reading to any reader as a live check.
+
+The comment above it was honest about the bypass: nine lines fail the pass and eight of those are
+correct writing. So the nine were adjudicated one at a time.
+
+* **Three are out of scope by construction** — a comparison against a colour the pack does not
+  ship (*"7.53:1 on pure white"*), a bound over a set of surfaces (*"11.8:1 on every tint"*,
+  *"1.97:1 on any of the six field stops"*), and a candidate measured in order to be rejected.
+  Named in `ARGUED` now, each with its line.
+* **Two were an indented table row misread as prose.** `line.startswith("|")` does not see a
+  leading space, so the closable-by-one-header-edit class under-reported and the prose residue
+  over-reported by exactly those two.
+* **Four were wrong writing.** The figure and its subject had drifted onto different lines, or the
+  token on the line was the *remedy* rather than the subject — and a reader cannot tell either from
+  a claim. All four were right about the arithmetic (recomputed) and are reflowed in `manpage` and
+  `roster` so the subject travels with the figure.
+
+**457 computed + 31 verified by the broad pass, 23 unguarded (4%), 6 prose the pass cannot pair.**
+The printed line reports the broad pass as its own class: counting a claim the run just verified as
+*unguarded* is the same untruth as the number it replaced, in the other direction. Two further
+corrections fell out of that — the total is a sum of buckets, so adding one without adding it to the
+total printed **495 claims where 526 exist**, a report getting quieter as coverage improved. And a
+palette plant had been asserting only that a partnerless prose claim *landed in a bucket*; it now
+asserts that a wrong number is **refused**, which is what the revived pass makes possible.
+
+### A carrier phrase was dropped twice, and the second time shipped in seven releases (B-052)
+
+`T1` gives an agent five skill descriptions and fourteen tasks and passes only at 0 misses,
+0 false loads. Its own record already carried the lesson: the 1.12.0 run scored 13/14 because
+*"scrubbed sections"* had been dropped to make room for a mobile trigger, and that phrase was
+the only thing in the description carrying **section by section**. It was restored, the set
+re-ran green, and the harness gained the rule that a description edit obliges the full trigger
+set.
+
+**The rule did not hold.** `v1.37.5` (`de09f9e`, *"an unqualified landing page reaches both
+crafts"*) removed `scrubbed sections` while the description **grew** from 1009 to 1021 chars —
+additive by every appearance. T1 was not re-run, so the regression shipped in **v1.37.5 through
+v1.45.0**, seven releases. It was found by walking the phrase across every tag, not by reading
+a diff, which is the only reason it was found at all.
+
+Restored (1007/1024), and the class is mechanical now: `validate_t1_carriers_survive` pins the
+twelve phrases T1's tasks depend on to the description and fails when one goes, watched failing
+against the v1.37.5 edit itself.
+
+**What this does not do is stated where it matters.** A string test cannot prove a phrasing
+routes here rather than to `super-ux` or `copywriting`. The last fresh-context T1 result is
+2026-08-11 and the description has changed six times since, so T1's record now says *do not read
+the green below as current* directly above it. That measurement is what B-052 stays open for.
+
+### Every plant names the check it exercises, and one of them never had (B-051)
+
+The self-test's fourth tuple element is the string the output must contain — the only thing
+separating *the gate went red* from *the check I wrote went red*. **13 of 60 plants carried
+none.** Running those thirteen one at a time and reading what each actually made fail turned up
+the defect the mechanism exists for: the container plant replaced a bullet's **label**
+(*"Container queries"* → *"Breakpoints"*) while `scoreboard` answers the container question
+twice in the same section, so `CONTAINER_ANSWER` still matched, the check stayed correctly
+green, and the `.cursor` mirror-drift check supplied the red. **It had reported `caught` on
+every run since it was written.** Four more of the thirteen produced a mirror failure ahead of
+their own, which is the same defect one reordering away.
+
+All thirteen carry their own check's string now, the container plant removes *every* answer in
+the section rather than one, and `validate_every_plant_names_its_check` makes `expect`
+mandatory — with a plant that strips one, planted into the plant table itself.
+
+### The run-stamp count is recomputed, and the retirement window is readable (B-051)
+
+The Run stamps preamble said **twenty-six** rows carry no `Diverged?` answer; the table held
+**thirty-four**. In the one document whose subject is what a run may conclude, that is the same
+restated-number defect this gate refuses everywhere else — and it matters because the third
+retirement trigger reads a stamp *count*, so the number decides what may be retired on absences
+of evidence. `validate_reconstructed_stamp_count` recomputes it and fails when the paragraph
+disagrees.
+
+A derivation was tried before any of this and failed honestly: `Diverged?` cannot be recovered
+from the Log, because **0 of the 34** rows have their commit named there. So the row's own
+prescription was executed instead — this session's five runs wrote their own stamps. The newest
+five now answer, which **opens the third trigger for the first time in eighteen releases**.
+`validate_retirement_window` prints OPEN or CLOSED, and the trigger's wording now states that an
+instruction may only be retired when all five newest stamps answer.
+
+Two bugs in the new checks, both caught by their own output. `\s*$` under `re.M` eats the
+newline, so `finditer` returned **every other row** — 29 of 57 — while the check printed a
+confident number over half a table. And `"five rows"` matched the preamble's own aside (*"a list
+that is short by five rows"*), so the count check read 5 against 34 and failed for the wrong
+reason.
+
+### The doctrine's duration bands reach the token layer (B-049)
+
+`validate_motion_bands` reads a pack's `## Components`, `## Micro-interactions` and
+`## Motion flavor`, because prose is where a pack prescribes an interaction. It never read
+the token layer, and the board named the hole with its instance: `bulletin`'s
+`--dur-panel: 0.4s; /* the nav dropdown */`, 150 ms past §3's dropdown band and invisible to
+every check here.
+
+`validate_token_comments_respect_the_bands` reads a **declaration** whose token name or
+trailing comment names a banded element. **9 subjects across 8 packs, all inside their band.**
+
+The 30 durations mentioned inside comment *prose* were extracted and deliberately left out.
+Every one of them describes the reference (*"the reference declares `--t-normal: 250ms`"*),
+quotes the doctrine's own band, or explains a reduced-motion collapse. A check reading those
+flags a pack for correctly reporting what it measured, which is how a gate teaches an agent to
+route around it.
+
+**Adjudicating the nine found what the row had not.** The same token was called *the nav
+dropdown* in the CSS and *the nav sheet* in the pack's own Motion prose, and §3 bands dropdowns
+at 150–250 ms and sheets at 200–500 — one word made 0.4s a violation and the other made it
+correct. It is a sheet: full-width, `--dropdown-min-h` 420, a `--surface` plane with a bottom
+`--line`. **The naming moved and the value did not.** 0.4s is a pack choice rather than a
+measurement — the reference's whole budget is an entrance fade, a 0.12s press and a 0.3s hover
+— and changing shipped motion on the strength of a word in a comment is the wrong repair. What
+a pack's two files disagreeing costs any band gate is filed as B-122.
+
+The first version of the band table silently judged **eight** subjects where nine exist:
+`\bdropdown\b` does not match *"dropdowns"*, so `awning`'s `--dur-control: 200ms /* DERIVED —
+dropdowns and selects */` escaped it. That value is legal, so nothing was hidden — the ninth
+subject would have been the one to hide an illegal value written in the plural.
+
+### Two scoreboard ratios described pairings that never render (B-047)
+
+All 30 ratio claims in `scoreboard` were recomputed. **28 agree** to within the repository's
+0.1 tolerance — including the composited focus halo, which measures 1.28 against a stated 1.29.
+The row predicted five disagreements; there were two.
+
+A Gotcha warned that the reference's positive delta is unreadable and put the number on the
+wrong field: `#00D492` **on white — 1.84:1**. 1.84 is its ratio against this pack's paper
+`--bg` `#FAF9F5`; on true white it is **1.94**. A reader who checked the figure against white
+got a different number and a document that looked wrong about the very thing it was warning
+them about.
+
+The ring table's header measured every row against `--bg`, and the `--ring-focus-sand` row
+confines its token to `--surface-sand` — the second ring exists *because* the first misses the
+floor there, so `--bg` is the one surface it cannot be about. Its stated **15.88** was correct
+arithmetic about a pairing that never renders; on sand the token measures **14.60**. The table
+has a `Measured on` column now.
+
+Both were fixed the way the row asked for rather than by editing a digit: each figure names a
+partner the token layer can pair, so the palette gate **computes** them on every run.
+`computed_at_least` moves 454 → 456. An unguarded correct number is one edit from an unguarded
+wrong one.
+
+**The broad check was tried and rejected.** 58 rows across 24 packs name a surface other than
+their header's base, and in nearly all of them the row's *subject* is a surface token, measured
+correctly on the declared base. A gate that flags those is a gate an agent learns to route
+around. Only a row that **confines** its token elsewhere is making the mistake, and after the
+fix there are zero of those — so `validate_confined_tokens_measured_where_used` ships as a
+regression guard with a plant, and says `0` out loud rather than implying a measurement. The
+undecidable half is filed as B-121.
+
+Two footnotes on the row itself. Its own coverage figure — *150 unguarded at 22 packs* — had
+gone stale during the same day and reads **57 unguarded (11%) at 17 packs** now. And the plant
+for the new check needed correcting: changing only the table row left a 4-cell row under a
+5-column header, which the table-shape guard catches first, so the plant was proving that guard
+worked. The defect as it shipped was internally consistent, and the plant is now too.
+
+### Every duration answers the reduced-motion query, and a promise is checked in the kit (B-045)
+
+`validate_reduced_motion` was a floor: a token layer had to *have* a branch and the branch
+had to collapse *something*. A layer collapsing one duration of nine passed it.
+`validate_every_duration_answers_reduce` asks the whole question — every duration declared
+outside the branch appears inside it, collapsed to an instant or re-declared at its own value
+**with a reason** — and it verifies the part the board row asked for by name: where the reason
+promises a stop in the component layer, the kit is checked for that stop.
+
+Now **29 layers, 0 silent durations, 6 kept with a reason, 2 promising a component-layer stop
+and both keeping it.** `0.01ms` counts as a collapse rather than an exception: it is the
+standard idiom for *instant but still fires `animationend`*, and of the 15 in-branch declarations that are not plain `0s`, **9 are that idiom and 6 sit above an instant**. Both figures are computed from the tree, not restated: the pair this paragraph carried until 2026-08-21 was *8 of 11*, and neither number resolved against any reading of *re-declaration* the document defined — the product tier of the certification measured it
+re-declarations are exactly that.
+
+**Two layers were fixed.** `prism` had added `--dur-press: 0.14s` and never updated its branch.
+`atrium`'s four `--flute-dur-*` are kept on purpose — under reduce the refraction canvas is
+*removed* and the still underneath becomes the hero, so a collapsed duration would state the
+opposite of what happens — and that reason now lives in the CSS beside the values instead of
+only in prose a gate cannot read. The row's own measurement was half stale: the uncollapsed
+pair was `atrium` and `prism`, not `roster`, which already used the re-declare-with-reason form
+this check now requires.
+
+**The first version of the check reproduced the defect the row exists for.** It returned early
+on any duration that collapsed, so `pigeonhole`'s `--dur-marquee` — which goes to `0.01ms`
+precisely *because* the kit pauses it, since 0.01ms strobes an infinite animation at the reader
+the query protects — never reached the kit verification. Its plant stayed green. The promise is
+a property of the layer, not of one token's verdict, and is checked once per layer now.
+
+### A gate constant was shadowed, and Python said nothing (B-120)
+
+The check above shipped broken and passing. A new `DUR_DECL` near line 3120 of
+`test/validate.py` was overwritten by an existing module-level constant of the same name at
+3716, whose capture groups mean something else — so the new check parsed every duration with
+the wrong regex and printed **`0 kept`** against a tree where six were measured by hand. It
+reported a number instead of failing, which is the exact shape this repository keeps finding.
+
+`validate_gate_has_no_shadowed_names` walks the three gate modules with `ast` and refuses any
+top-level name bound twice. It found one pre-existing duplicate — `CSS_COMMENT`, defined twice
+identically, harmless and now deduped. Its own plant needed two fixes to prove anything: the
+check first read `Path(__file__).parent`, which is the one directory a plant can never reach
+because the self-test runs the real module against a copied tree.
+
+### A pack cannot prescribe a token its own layer never defines (B-043)
+
+`instrument-console`'s prose named `--dur-press` and its token layer never declared it.
+`var(--dur-press)` is then an undefined custom property — invalid at computed-value time,
+falling back with no error anywhere — which is board B-038's class, described in a comment
+inside the very file that had just acquired a fresh one. The commit that introduced it was
+the one closing B-042, hours earlier.
+
+`validate_prescribed_tokens_resolve` refuses it, and the naive form of that check was
+measured before it was written rather than after: **19 of 29 packs quote a token their layer
+does not define**, and nearly all of it is a pack quoting its *reference's* vocabulary —
+`datasheet`'s `--gray-6`, `ora`'s `--font-lora`. What separates a promise from a quotation is
+the prefix family: a pack naming `--dur-press` while defining four other `--dur-*` is speaking
+its own vocabulary. That cut 60 candidates to 11; excluding the `--font-` family on principle
+(a pack layer deliberately does not own font loading) left 3; one is a declared exemption with
+its reason and a line reference. Two were real — the second, `roster`'s `--pattern-grid`, is
+fixed in the prose rather than by inventing a token, because the pattern is composed from the
+five `--pattern-*` values the layer does define.
+
+**The plant found a hole in the check itself.** Dropping `--dur-press` from `:root` left the
+reduced-motion block's copy standing, so the token still counted as defined. A declaration
+inside `@media` is not a definition for anyone not asking for reduced motion; `_outside_media`
+now strips those blocks. Zero tokens sit in that position today across 29 layers, which is why
+it was worth closing before the first one does.
+
+**`--section` was the library's only raw `vh`.** `clamp(9rem, 24vh, 20rem)` drove section
+padding off the viewport, re-laying out every section when a mobile address bar shows or hides.
+It is `24svh` now, with the reasoning written where the five packs carrying a `--hero-min-h`
+wrote theirs: a hero wants the viewport visible right now and takes `dvh`; padding wants a
+height that does not move.
+
+Two of B-043's five claims were false and the row says so. The layer is **4869 bytes, 7th
+smallest of 29** — not 1.4 KB and not the smallest; `briefing-room.css` is smaller. And the
+focus-ring claim misread template rule 5, which governs fallback order for derived colour
+rather than requiring a ring token: **20 of 29 packs ship none.** The type-scale half is real,
+library-wide and refiled as B-119 with its measurement: 6 packs token it, 11 state the sizes in
+prose, 12 do neither.
+
+### One site-wide curve and a mandated scrub cannot both be total (B-042)
+
+`instrument-console` declared *"the one site-wide curve"* and then mandated scrubbed SVG
+instruments. `MOTION_DOCTRINE.md` §6 settles it — *under `scrub`, easing must be `none`;
+the scrollbar is already the clock* — and easing a scrubbed timeline eases against the
+scroll position twice, which the doctrine itself calls the family's most common motion
+bug. The pack never carved the exception, so an implementer reading it alone landed on
+exactly that bug. **The carve is now written into its Motion tokens**, and
+`validate_scrub_carves_its_exception` refuses a pack that declares one curve, mandates a
+scrub and stays silent. Measuring the shape across the library found **five** such packs;
+the other four are pinned at `scrub_uncarved_at_most: 4` and filed as B-118, because each
+carve has to be written against that pack's own motion section rather than swept.
+
+Two more things the row was right about. The pack's `Contract: core` line said the hero,
+its line ceiling and per-component states were *"yours to decide"* while the pack states
+all three — **the declination is narrowed to what it actually leaves open**, because a
+reader who believed the old wording would have decided them twice. And its fastest step,
+`0.18s`, sits outside the doctrine's 100–160 ms button-press band with no deviation clause
+to appeal to; `--dur-press: 0.16s` is the band's own ceiling, derived from the doctrine
+rather than picked, since ADR 0003 forbids widening this pack by invention.
+
+The first plant for the new check was reported *caught* with the ratchet lifted past the
+defect — an edit to a pack trips the `.cursor` mirror check too, so the plant had been
+proving the mirror check worked and saying nothing about the check it was written for. It
+carries the required `expect` string now: with the pin at 5 it reports `MISSED`.
+
+### The library's contrast coverage was 15 claims smaller than it said (B-013)
+
+`stated ratios: … 336 guarded` counted a claim as covered before computing a single
+pair. Fifteen of those 336 had never been through the arithmetic — a claim whose only
+named partner is itself (`--ink` in a table declaring `--ink` as its base), or whose
+subject is a composite the gate cannot resolve to a solid (`ledger.css:33`'s `--muted`).
+**The word `guarded` is gone from the line.** What prints now is what happened:
+`513 claims — 348 computed, 15 named a partner this pack cannot pair, 150 unguarded`,
+and each of the 15 is listed by `file:line` on every run, because a number nobody can
+act on is a number nobody acts on.
+
+**The unguarded set is classified instead of lumped.** 177 said nothing about which of
+them a person could close. The split says it: `0 in a table declaring no base, 10
+placing the ratio at a gradient stop, 140 prose`. The 27 that sat in five tables closed
+by declaring the base each table's own numbers already resolve against — `atrium`,
+`router` and `orchard` on `--bg`, `cyclorama` on `--ink` with four cells naming
+`--field-2`, `prism`'s components row on `--accent`. No ratio was changed and no base
+was chosen: each was found by computing which token produces the numbers already
+written. `atrium`'s row also had to spell `--ink-2` and `--ink-3` instead of `-2` / `-3`,
+because a shorthand is not a token reference.
+
+**A third guard was written and thrown away, and that is the useful half.** A wrapped
+comment leaves the partner alone with the number (`4.40:1 on --bg-deep.`), so reading
+the subject off the line above looked like free coverage — a probe said it resolved 7 of
+the 15. It did not: the probe accepted a match under *either* comparison mode, so any
+ratio above the claim counted as agreement. Inside the gate it produced two false
+failures at `ledger.css:33`, where the real subject is a composite and the line above
+offers a different token. Substituting a subject is the same act as guessing a partner,
+which is what killed attempts one and two. The reason is recorded where the code would
+have gone.
+
+### A pack's own semantic tokens were compared against nothing (B-017)
+
+`validate_theme`'s peer set is `STATUS_TOKENS` plus `--accent`/`--primary`/`--cta`, so a
+pack carrying meaning in tokens of its own naming was checked by nothing. `field-notes`
+exists to render provenance, and all four of its provenance tokens were excluded **by
+name**: `--brand-ink` (`[INFERRED]`) and `--witness-ink` (`[AMBIGUOUS]`) are **3.2 apart**
+at full colour and **0.8 under deuteranopia** — the two labels a reader most needs to tell
+apart, on the pack whose reason to exist is provenance. The dark theme is worse than the
+board recorded: `--witness` and `--danger` are **1.2 apart** there, 0.4 under deuteranopia.
+
+**The peer group is now whatever the pack declares**, and none of the three forms is new
+syntax — all three already ship:
+
+- the state map in a token layer, `[EXTRACTED] → --verify-ink · [INFERRED] → --brand-ink`;
+- a distinctness claim, *"destructive — kept distinct from `--witness`"*;
+- the pack's own disclosure that two tokens share a hue family.
+
+What it deliberately does **not** do is infer a set from token names. `--brand` and
+`--brand-ink` are one colour's two jobs; `--witness` and `--danger` are two meanings that
+share a hue. Nothing in the naming separates those cases, so the pack has to say which it
+means.
+
+**A disclosure carries its own arithmetic.** *"One hue family, 2.8 apart"* is an honest
+answer only while 2.8 is what the values produce, and `validate_stated_ratios` cannot see
+it — that check reads contrast ratios and this is an OKLab distance. Any distance a
+disclosure states is now computed, bound to the quantity it names: `N apart`, `N at full
+colour`, `N under deuteranopia`.
+
+**No hex moved.** `Origin: graphify.com (2026)` makes these values measurements, and a hue
+read off a reference may not be re-stepped. So `field-notes` answers in writing: rust and
+red-orange are one hue family by measurement, the provenance state is carried by the
+bracketed word and its border, and `--danger` is separated from `--witness` by form — a
+filled control, never an inline tag. The sentence *"kept distinct from `--witness`"* was
+false at 2.8 apart and is gone; the Bans now say the conflation they forbid is **in role**,
+and a new ban forbids a provenance state carried by colour alone. `Three semantic hues and
+no fourth` became `three semantic ROLES, and two hue families rather than three`.
+
+**Four false-positive shapes were found by measurement and closed, and each is recorded
+where the code is.** A one-tier floor called a pair 19.7 apart at full colour a hard
+failure. A substring test let a line about `--witness-ink` excuse `--witness`. A
+paragraph-wide bind gave every figure in `cyclorama`'s status bullet to every pair — six
+failures against six correct sentences. And two separate regexes cut *"6.4 apart under
+deuteranopia"* in half, comparing a deuteranopia figure with a full-colour distance.
+
+### A row that states an instruction instead of a ratio was checked by nothing (B-048)
+
+`validate_stated_ratios` reads a number the document states. A Components row often states
+none — it states an **instruction**: *"`--surface-2` fill at 4% ink, `--ink` at 50%"* — and
+nothing composited it. `scoreboard`'s secondary button shipped a **10–12px/600 label at
+3.16:1** that way, below AA, with every gate green over it because there was no number to
+check.
+
+**One sub-claim of the board row was itself wrong**, and the measurement says so: the row
+held that 4% ink over the paper *"resolves lighter than the declared `--surface-2`"*. It
+resolves **darker** — luminance 0.873 against 0.941 — which is precisely why it pushes the
+label below AA rather than above it.
+
+The fix replaces two ad-hoc composites with tokens the pack already ships (`--surface-2`
+fill, `--ink-soft` label) **and states the ratio it renders**, 4.55:1. The first version of
+the fix did not: correcting the spec moved the pair out of every check, which is the same
+trap as B-017, where a false distinctness claim was the only thing making a pair
+observable. A correction that removes the claim removes the coverage with it.
+
+**Three behaviours with no token now have one.** `--t-display-sm` 42px and `--t-display-md`
+44px carry the headline's two unnamed breakpoint steps; `--label-col` 95px,
+`--label-col-narrow` 75px and `--numeral-col-narrow` 70px carry the ledger's geometry, which
+the kit had as three `width:` literals and the pack as prose numbers. Every value is the
+measured one that was already written down — a token names it, it does not choose it.
+
+### A pair a Components row prescribes was computed by nobody (B-050)
+
+A ratio is checked when the document states one. A Components row usually states none: it
+**prescribes a pair** — *"`--accent` fill, `--on-accent` label"* — and its hover and active
+cells replace one half of it. `contrast(label, fill)` is derivable at every state of every
+component, and nothing derived it. That is how `instrument-console` shipped a pressed label
+at 3.06:1 (B-039): the press dimmed the fill and the label stayed.
+
+`validate_prescribed_pairs()` reads the resting cell for a fill and a label, applies each
+state cell's `fill →` / `label →` swaps, and holds the result to the AA floor for the row's
+type size. 13 rows across 7 packs name both halves and carry 8 state swaps between them.
+
+**Both defects it found sit in the two rows the widened phrasings added:**
+
+- **`maquette`'s leader label is invisible.** The row prescribes `--ink` on a `--model-top`
+  fill, and those are the same hex — `#FFF9F4`, **1.00:1** — in the pack whose central rule
+  is that type and object are built from one cream. The pack now ships `--on-model`
+  (`#151515`, the field's own colour) at 17.49:1, and the token layer says why in the place
+  a reader hits it.
+- **`prism` sets a `$` prompt in `--ink-faint`** at **2.31:1**, a token its own layer calls
+  *"furniture only, never text"*. It is `--ink-soft` at 4.70:1 now — the token that layer
+  says *"MAY carry body copy"*.
+
+As the board row predicted, this check could **not** have caught B-039: `instrument-console`
+is a core-contract pack with no Components table, which is also why the defect survived
+there. 16 rows state half a pair and the count prints on every run (B-059).
+
+### 88 prose ratios joined the arithmetic, and four of them were wrong to (B-003, B-053)
+
+140 stated ratios sat in prose that named no partner, so nothing computed them. They were
+not closed by widening the check — a heuristic that guesses the other side is what this
+file has thrown away three guards for. They were closed by testing **one hypothesis**
+against all 140: *this ratio is against the pack's own field.* **88 agreed with it
+arithmetically**, claim by claim, on 77 lines.
+
+Those lines now name `--bg`. Many already said *"on the field"* or *"on the paper"* — the
+same fact, written so a machine cannot read it — and the rest said nothing. Coverage went
+from `353 computed / 150 unguarded` to **`448 / 57`**, 29% unguarded to 11%.
+
+**Four of the 77 went red on the first run and every one was a real find**, because giving
+a line a partner puts its *other* claims under the arithmetic too:
+
+- `blueprint`'s `3:1` is the non-text **floor**, and the word *floor* had wrapped to the
+  next line, so `RATIO_SKIP` could not see it. Reflowed.
+- `proscenium`'s subject was the words *"Body ink"* — no token — and its *"cool act"* is
+  `--panel-2`, which the token layer confirms at **16.49:1**. Both named now.
+- `tenor`'s subject sat on the line above its claim, the same shape as the fifteen the
+  palette run prints as `unpairable`.
+
+The 52 that refused the hypothesis were left alone and stay on the board, because a claim
+whose partner is a panel, a literal the pack does not ship, or a token on the previous line
+needs its own answer, not a rule.
+
+### The board's own table hid two `high` rows for ten iterations
+
+A row's priority and status are read from the column its header names. An unescaped `|`
+inside a row shifts every cell after it, and three rows here carried a shell pipe inside a
+code span — `sort -u`, `cut -d/ -f2`, `grep … | wc -l`. Their status cells came back holding
+`sort -u`, `` `datasheet` run, stage 0 `` and `medium`, so the queue that orders this
+repository's work never saw **B-004** or **B-014**, the only two `high` rows in the file.
+
+The rows were never wrong. The table was, and nothing looked at the table.
+
+`validate_board_columns()` now takes each table's width from its own `| id | … |` header and
+refuses any row that does not match, across `backlog.md` and `verification.md`. Planted with
+an unescaped pipe and watched refusing. Floor 3560 → 3619, one check per row.
+
+Measured across the whole family in the same pass: **13 such rows in five repositories**,
+including rows whose columns were simply missing or doubled. All are fixed and the count is
+zero.
+
+### A tag can no longer publish a release no clone can reach (B-014)
+
+A tag pushed without its branch is a release that exists on npm and on GitHub and nowhere a
+`git clone` can find it. Nothing fails loudly, because the release workflow fires on the
+**tag**: the release is created, the package publishes, and `origin/main` still reads the
+previous era. This repository hit it twice — the `v1.4.0` ghost on 2026-08-04, and `v1.18.0`
+on 2026-08-12 with `main` two commits ahead of `origin/main` while the tag was public.
+
+The release job now refuses it: `git merge-base --is-ancestor HEAD "origin/$DEFAULT"`,
+against the repository's **actual** default branch read from the API rather than a hardcoded
+`main`, so a fork that renamed its branch is not told its correct tag is unreachable. The tag
+checkout gains `fetch-depth: 0`, because ancestry on a shallow clone is not an answer — it is
+whatever the depth happened to give.
+
+Watched refusing against a reproduction of the defect: a commit that was never pushed,
+tagged, and the ancestry test says so. Watched passing on this repository's real HEAD. **All
+eight members carry it**, and the umbrella requires both halves in each, watched failing when
+either was removed.
+
+### `atrium` is widened, from the reference rather than from its own tokens (B-004)
+
+Three packs were held at `Contract: core` because widening them from the token layer would
+have been *inventing values with a citation attached* — the one thing the pack layer exists
+to prevent. Their references are addressable, so the honest route was always a measurement
+pass, and this is the first of three: `atrium`, read on 2026-08-20 from
+`functionhealth.com`'s own `/_astro/utilities.*.css` and `/_astro/PageSections.*.css`, mapped
+through the reference's **own** token names — `--fill-04` is `--accent`, `--fill-03` is
+`--ink`, `--fill-01` is `--bg`, `--stroke-02` is `--line`.
+
+- **`## Components`** is the reference's single `.fh-button` class and its four skins, with
+  every state: the pill radius (`999px`, not `--radius-lg`), 48px and 40px heights, the exact
+  paddings, and the behaviour that matters — **the press does not dim the accent, it replaces
+  it with the ink**, so the label goes from 4.59:1 to 13.47:1 rather than losing contrast.
+  Hover and focus-visible share a fill, so the 2px accent-at-35% ring is the only thing that
+  separates keyboard from pointer.
+- **`## Hero`** is the numbers of the full-viewport media hero: a 540px text column at the
+  bottom of the stage, `--h1` fluid from **57px to 80px** at weight 300 and `line-height: .9`,
+  the `blur(15px)` eyebrow pill over 7% white, a stat rail with 58px hairline dividers, and
+  what changes below 768px — the stage becomes 666px and the rail leaves the media for a
+  `--bg` bar beneath it.
+- **`## Responsive`** is its **184 media queries, counted**: 768px and 1100px carry 140 of
+  them. It also records the thing a reader would otherwise get wrong — **type does not step at
+  those boundaries, it slides**, `clamp()`-ed between 376px and 1440px, so mixing a fluid
+  heading with a font-size media query breaks the rhythm. And it answers the container
+  question with a technique rather than a shrug: the reference declares `container-type` in
+  exactly one place, a stage whose every dimension is in `em` and whose `font-size` is driven
+  by the container width, so one number scales the whole composition. 152 `vw` expressions
+  against 2 container-relative ones.
+- **`## Signature element`** names the fluted-glass hero that Signature motifs already
+  measured, and says why it is the signature rather than one motif of several.
+
+**The pass found a value the reference does not ship.** `--accent-tint` was `#F7EEEB` —
+which appears **nowhere** in the reference's CSS or its HTML — against its
+`--orange-10p: #F6E9DC`. The mapping is not a guess: its neighbour `--green-10p` is `#f1f3e5`,
+this pack's `--good-tint` to the character. Corrected, with the search that found it recorded
+in the token layer.
+
+It also found what no section had ever said: **`--accent` is 4.17:1 on `--surface`**, below
+AA. The accent is safe as a label on `--bg` and on `--ink`, and not on the cream card. The
+reference never puts one there; now the pack says so.
+
+Contract split **7 core / 22 widened → 6 / 23**. `orchard` (gutgutgoose.com) and
+`editorial-luxury` (prowl.chat) remain, one measurement pass each.
+
+### `orchard` was measured and deliberately not widened (B-004, B-110)
+
+The second of the three packs held at `Contract: core`. Its reference is addressable, so the
+same measurement pass was run against `gutgutgoose.com` — and it stopped, on purpose, at
+`## Components`.
+
+**What the pass established.** All **13** of `orchard`'s token values are present in the
+served document — 10 as hex and 3 as `rgb()` triples, which a hex-only search had first
+reported as missing, so the first answer was wrong and looking harder was the fix rather than
+concluding. Its responsive model is the **opposite of `atrium`'s**: three fixed frames
+(desktop ≥1200, tablet 810–1199.98, phone ≤809.98) with a **390px phone frame** carrying a
+342px column, and the tablet/phone tier changing 61 rules dominated by `width` (32), `order`
+(22), `height` (16) and `padding` (14). **No `clamp()`, no container query, no `vw` at all** —
+variants are swapped, not scaled. The slab radii are 12px (23 uses), 20px (10), 24px (9) and
+40px (4).
+
+**Why it was not widened.** `## Components` is per-component states, and a Framer document
+serves none. The contract is all-or-nothing by design — *touch one of the four and you owe
+all four* — so writing three from measurement and the fourth from the token layer would be
+precisely the failure `core` exists to prevent.
+
+**And the pass found a contradiction worth more than the widening.** The pack calls its candy
+pill *"the only orange object on the page"* and `--cta` *"the ONE action colour, a fill"*.
+On the served page `#FA7241` occurs **only inside SVG `fill=` attributes** and **never as a
+`background-color`**; what paints controls there is `--surface`, `--bg`, **`--primary`
+`#78934a` (14 declarations)**, its 10% wash, and `--surface-ink`. Three readings fit the
+evidence and it does not choose between them, so the pack was **not** rewritten — filed as
+B-110, to be settled in the same rendered pass that `## Components` needs.
+
+### `## Hero` had two readings, one worked example, and no check (B-007)
+
+The board said the standalone reading was missing from the library. **It was not** —
+`ledger` had already written it: *"A console's hero is its first viewport after sign-in"*, with
+a kicker, a title at a console size rather than a display size, the answer before the evidence
+that explains it, and an explicit list of what the screen must not carry. What was missing is
+that **the skeleton did not carry it**, so every console-pack author re-derived the heading
+alone, which is exactly what the T7 author did and said.
+
+The skeleton now names both readings with a worked example each — `datasheet` for the
+marketing page, `ledger` for the console — and states that both owe the same two things.
+
+**And nothing had ever read the answer.** The skeleton has demanded a line ceiling *and the
+container width that keeps it there* since 1.5.0, and no check looked. Across the 23 packs
+that carry a `## Hero` **heading** — never the string, which a `core` pack names in its own
+contract line to decline it — there was exactly one live defect of each kind:
+
+- **`atrium` stated no ceiling.** Widened by this same loop one iteration earlier, with every
+  measurement in place and the one obligation the skeleton spells out left out. Fixed from the
+  reference: **two lines in the 540px column**, and the ceiling is a word budget rather than a
+  CSS rule — the reference's own headline is *"Check your health."*, three words.
+- **`showroom` stated a ceiling with nothing holding it.** Fixed from its reference, and the
+  answer is the interesting one: **no `max-width` holds it at all**. The headline sits in a
+  centred column with side padding only — 24px, 56px, **58px from 1024px** — and
+  `text-wrap: balance` at `line-height .95` does the rest, so the ceiling is a **four-word
+  budget**. The reference's headline is *"Welcome to agentic revenue."*
+
+Two plants, both against `maquette`, chosen by measurement rather than by eye: its hero carries
+exactly one phrase of each kind, so removing one leaves the other standing and each plant
+proves its own half. Floor 3619 → 3667.
+
+### Every pack now declares its two real silences, and the gate derives them (B-008)
+
+`Contract: core` named four sections and nothing else, so a pack's **other** silences stayed
+undeclared. The case on the board: `workbench` ships four status roles and no severity ramp,
+and a fresh-context agent building an incident list inferred a rank scale from role
+descriptions and discovered the gap by hitting it.
+
+The row offered two answers and **both were refused, on evidence**. A free *"also not
+specified here"* list is unbounded and uncheckable — every pack would write a different one
+and nothing could compare them. And four is not the boundary, because two silences are real
+and were counted before anything was chosen:
+
+- **Themes: 10 ship a dark twin, 13 ship one block, 6 ship a second block that is a
+  *surface* rather than a theme** — `[data-surface="dark"]`, `[data-surface="panel"]`,
+  `[data-state="alarm"]`. So *"does this pack have a dark mode?"* had three different answers
+  depending on how a reader looked, and `tenor` in particular ships something dark that is
+  not a dark theme.
+- **Rank: exactly one pack of twenty-nine ships an ordered ramp** — `tenor`, with
+  `--sev-ask` → `--sev-limit` → `--sev-never`. The other twenty-eight were silent.
+
+So each pack carries a `Themes:` line and a `Rank:` line, **derived from its own token
+layer**, and `validate_pack_declares_its_silences()` re-derives both on every run and refuses
+a declaration that has drifted from the thing it describes. That is the property a prose list
+could never have: the answer is checkable against the tokens, not merely present.
+
+Four plants — each line removed, and each claiming what its own tokens contradict. The
+skeleton carries both questions with the counts and the reason, and says plainly that a third
+question is only added when it can be answered the same way. Floor 3667 → 3783.
+
+The library's own counted-claims guard refused the first draft of that skeleton text, which
+said *"11 packs"* where the checkable phrase means the pack total. Reworded, not exempted.
+
+### The code graph is current again, and the refresh needed no key (B-009)
+
+The row had been re-filed as *staleness alone — a re-run, not a credential question*, at 13
+commits behind. By the time this loop reached it, the same loop had pushed it to **31**.
+
+`graphify update .` — the path whose own help says *no LLM needed* — moved `built_at_commit`
+from `de09f9e` to **`856f12c`, which is HEAD**. And it did not do the thing the row's history
+feared, where a code-only pass had once written fewer nodes than the curated graph held:
+**every class grew.** Documents **1797 → 2173**, code **1646 → 2105**, concept 243 → 340,
+rationale 126 → 155. In total **3812 → 4773 nodes** and 4185 → 4473 edges across 920
+communities. The shrink guard never fired because there was nothing to guard against — and
+the curated graph was copied to `graphify-out/2026-08-20/` before the run regardless, because
+that is the order the row itself insists on.
+
+**No placeholder labels survived.** 116 communities changed membership and were renamed by
+their hub; a scan of all 4773 nodes finds **zero** carrying a bare `Community N`.
+
+**Stage 9's own check passes.** All ten god-nodes resolve to documented seams — the two gates
+and their `check`/`read`/`main`, the pack skeleton, two style packs, the CHANGELOG, and the
+third-largest, `Component Spine`, which is a rationale node from `kits/atrium/src/Button.md`
+named in the CHANGELOG and in two specs.
+
+**Scope, stated rather than implied:** `graphify-out/` is gitignored — *machine state, not
+source* — so *current* means current on this machine. A clone still builds its own, and this
+close does not change that.
+
+### A count whose noun is only implied, and the one form that can be caught (B-016)
+
+`validate_counted_claims` reads *"N packs"*. It cannot read *"the accent role resolves to
+`--accent` in twenty-seven,"* where the noun is implied by context — which is how that one
+sentence sat wrong at ten, then thirteen, then fourteen, corrected by hand each time and
+unguarded again the moment it was.
+
+**Three patterns were measured and two thrown away**, which is the useful half:
+
+| Pattern | Spans | Verdict |
+|---|---|---|
+| every numeral with no counted noun after it | **2553** | ordinary prose — *three cards*, *one motion methodology* |
+| a numeral **closing its clause** | 62 | nearly all correct ellipsis — *one gate of four*, *wrong two times out of three* |
+| the same, in sentences naming a library noun | 69 | worse: mostly the word *one* |
+| **`the <numeral>` closing a clause** | **7** | precise — a definite article asserts a known set |
+
+Of those seven, three are library counts, and **one of the three was wrong**: `tenor.md:328`
+said token names are *"not uniform across the twenty"* while thirty-one packs ship. A live
+wrong number, invisible to every gate, found by narrowing rather than by widening.
+
+All seven now name their noun — which hands them to the existing check forever — and
+`validate_a_count_names_its_noun()` refuses the bare form. Both count checks read one shared
+corpus now; it was inline in one of them, and two copies of a source list are two lists that
+drift.
+
+**Naming the noun immediately exposed two more defects, which is the point of the exercise.**
+`the nine` became `the nine headings` and the gate refused it at once: the contract is
+thirteen. It reads `nine of the thirteen headings` now. And a *true subset* — the accent role
+resolving in 27 of 29 — is refused when written `twenty-seven packs`, because the check reads
+any `N packs` as a claim about the total. A subset is written **`N of the M packs`**, and that
+rule is now in the pack skeleton with the rest.
+
+The skeleton carries this as its **sixth** rule — and writing it corrected the section's own
+heading, which promised four rules over five. Floor 3783 → 3789.
+
+### `field-notes` ships thirty dark overrides that no prose carried (B-018)
+
+The pack declares `Themes: light + dark — a full theme twin` and its Palette documented only
+the light values. `tokens/field-notes.css:251-291` is now written out as a table, and writing
+it confirmed all three consequences the T24b run reported — plus one it missed:
+
+- **`--surface-2` and `--surface-3` are the same value** in dark, `#262019`, so the deepest
+  step the light table names does not exist there. **`--ink` and `--ink-2` collapse the same
+  way**, which the row did not catch.
+- **No `--dawn-*` stop is re-declared**, so `--hero-dawn` still ends `--dawn-7` `#E9ECDF` at
+  98% into `--bg` at 100%. In dark that is near-white meeting `#14110E` across two percent of
+  the gradient — the hard edge this pack's own Bans forbid. The Palette now says plainly:
+  **do not ship the hero in dark** until that decision is made.
+- **The pack's stated decision and its shipped theme disagree.** Gotchas says the forest
+  family was kept over the reference's warm-brown `.dark`; the layer ships `--bg: #14110e`,
+  which is that warm-brown. **The reference cannot settle it**: `graphify.com` serves no dark
+  theme at all today — no `[data-theme="dark"]`, no `.dark`, no `prefers-color-scheme`, and
+  none of the thirty dark values anywhere in the 488 KB it returns. Recorded where a reader
+  hits it and filed as B-112, not guessed at.
+
+The row's fourth claim was imprecise and is corrected rather than repeated: `SKILL.md` carries
+no list of packs with a mode collection. `FIGMA_BRIDGE.md` uses `workbench` as an *example*,
+and now says so, pointing at the `Themes:` line each pack carries.
+
+### A number this run published four hours ago was wrong
+
+The theme split shipped into the pack skeleton as **11 / 13 / 5**. The tree says **10 / 13 /
+6**. Two derivations of one number existed in the run that wrote it — a loose regex over the
+raw CSS, which counts any pack merely mentioning `dark`, and the strict one that generated the
+`Themes:` declarations — and the prose quoted the loose one.
+
+Corrected in all six places it had reached, and `validate_theme_split_is_derived()` now
+computes the split from the declarations and refuses a document that states another. Planted
+with the exact wrong number that shipped.
+
+### Routing: four of five findings held, and correcting one number found two more (B-019)
+
+- **The dial precedence rule had no third case.** *"The row that names the surface wins over
+  the row that names a mood"* does not arbitrate a row naming **neither** — *"trust-first:
+  regulated, public-sector, clinical"* names who reads it, not what they read it on. Both T24
+  branches hit exactly that and both split two-to-one on MOTION, which is the tell. The rule
+  now says an audience posture does not arbitrate DENSITY and **wins on MOTION and ornament**,
+  because that is the whole of what it constrains.
+- **`widened` appeared zero times in the selection text**, although every non-core pack
+  declares it. It appears now, beside the core marks.
+- **"Fetch the kit first" was unconditional inside a core-scoped paragraph.** A kit exists for
+  every pack; for a widened pack the kit and `## Components` must agree, and for a core pack
+  the kit is the only answer either of them gives.
+- **The greenfield contradiction does not hold against the current text.** Lines 219–226 give
+  the path explicitly — *build the page in the pack the brief's register points at*, with
+  `workbench` and `showroom` as the no-signal defaults. Recorded as not-reproduced rather than
+  repeated.
+
+**Correcting one count exposed two more in the same file.** `SKILL.md` said **seven** packs are
+on the core contract where six are; it still marked `atrium` `**core contract**` **eight
+iterations after this loop widened it**; and it said **twenty-two** widened where twenty-three
+are — a remainder the gate's own message records as stale across two consecutive releases.
+
+The check that existed compared the prose *count* with the table, so it could only fire when
+somebody edited the prose. `validate_table_marks_match_the_contract()` now reads the table
+against each pack's own `Contract:` line, **in both directions**: a mark left behind after a
+widening, and a core pack the table forgets to mark. Planted both ways.
+
+Finding (1) is real and larger than the row said — filed as B-114 with the spread measured:
+the motion ceiling has **two homes**, the doctrine names ten packs, four state one in their own
+Register, the overlap is two, and six packs marked `(standalone)` have a ceiling in neither
+place.
+
+### The motion ceiling had two homes; now it has one (B-114)
+
+Measured before anything moved: `MOTION_DOCTRINE.md`'s standalone passage named ceilings for
+**ten** packs, **four** packs stated one in their own `## Register`, the overlap was **two**,
+`awning` stated a 4 the doctrine never mentioned, and **six** packs marked `(standalone)` in
+`SKILL.md` had a ceiling in neither place. A reader who chose off the table and stopped took a
+ceiling that might not exist.
+
+**The Register is the home**, because it is what a reader of one pack actually has. Eleven
+Registers gained a statement — each relocated from the doctrine with the doctrine's own reason
+(`workbench` and `briefing-room` at 3 as not cinematic; `ledger` at 3 with its three state
+loops stopped under reduced motion; `pigeonhole` at 4 at the family's floor; `roster` at 4,
+keeping only its sticky nav; `field-notes` with **no hard ceiling**, the one standalone
+exception) — or, for the six the doctrine never named, from the default it does state: the dial
+is cut by §1's frequency table alone. All **fifteen** standalone packs now state a ceiling or
+state they have none.
+
+`validate_motion_ceiling_has_one_home()` asks three things: every `(standalone)` pack states
+one, no Register states two, and where the doctrine names a number the Register agrees.
+
+**Two defects in the check were found by its own plants**, which is why they exist. Reading the
+number *before* the "no ceiling is pinned" phrase made a Register that says both report the
+number and pass — hiding the exact case the next block was written to catch. And the doctrine
+writes a ceiling two ways, *"`pigeonhole` … above **4**"* and *"`ora` at **4**, `tenor` at
+**4**"*, so reading only the first left four of the packs it names uncompared.
+
+Corrected, it failed immediately on `proscenium` — where **this run's own default sentence sat
+on top of a Register that already named 4**. Removed. Floor 3795 → 3830.
+
+### The two radius rules were never competing (B-020)
+
+The board said the skeleton teaches a rule a shipped pack records as a defect, and asked which
+one is right. **Both are** — they govern different layers, and the row's framing was the thing
+to fix. Measured across the library: **sixteen packs state radius arithmetic.**
+
+- **The proportional ramp defines the token set.** `--radius-sm: calc(var(--radius) * 0.6)`.
+  That is where a value comes from.
+- **Subtraction adjusts a nested instance** where the padding is known:
+  `calc(var(--radius-control) - 4px)`, and `datasheet`'s 16 − 8 = 8 is subtraction done right.
+
+`ora` uses both and says which is which. `field-notes`' recorded error was **conflating them**:
+its rule said subtraction, its worked example subtracted 12 from 12 and announced 7.2, and its
+token layer used neither — three systems, and an agent applies the one it can read.
+
+The skeleton teaches both layers now, what each governs, and that failure by name. And
+`field-notes`' correction no longer over-reaches: it said *"not by subtraction"*, which reads as
+*subtraction is wrong* rather than *subtraction is the other layer*.
+
+`validate_worked_radius_sums_compute()` refuses a written subtraction that does not compute,
+excluding a `*(Corrected …)*` note because that is a record of an error rather than a claim. It
+ships with **no live subject** — every sum in the sixteen packs computes — so it ships with a
+plant instead. It then caught **this run's own cautionary quote in the skeleton**, where the
+broken sum had been reproduced verbatim: a reader copies what they see, so the warning is now
+written as prose rather than as a copyable equation. Floor 3830 → 3833.
+
+### The Components list is a contract, and the library had already decided (B-025)
+
+The board asked whether the skeleton's six component classes are a contract that ten widened
+packs owe entries for, or a menu an author picks from. **The library's own behaviour settles
+it.** Counted over the twenty-three packs carrying the heading: **buttons in all 23, cards in
+22, inputs and navigation in 21, loaders and empty states in 19.** Calling the list a menu
+would license the gaps rather than close them.
+
+So the skeleton says **contract** — and says the thing that makes it liveable: **"none, and
+here is why" is a full answer.** What is not an answer is silence, because a reader cannot tell
+a pack that has no loader from a pack whose author ran out of afternoon.
+
+**`atrium`'s four gaps are closed by measurement**, read again off `functionhealth.com`:
+
+- **Inputs** at the button's own 48px control height, a pill border in `--line-strong`, a white
+  fill rather than `--bg`, an `--ink-3` placeholder — and the honest note that label position
+  and error placement are **not on the reference's page**, so they are the implementer's.
+- **Navigation**'s mobile shape in full: the panel's fill goes `--accent` closed → `--ink`
+  open, with an `::after` strip carrying the same fill so the two cannot disagree mid-transition.
+- **Empty states** as `--ink-2` at 16px/300, centred, no illustration and no button.
+- **Loaders: none** — and that is measured, not omitted. Zero skeleton, shimmer or spinner
+  rules across both stylesheets.
+
+`validate_component_classes_are_answered()` pins the remaining **nine** (class, pack) pairs as
+a ratchet that may only fall. It is a ratchet rather than a sweep because each fall needs its
+own reference read, and inventing an answer is precisely what the pack layer exists to prevent
+(B-115).
+
+**Its plant needed two rewrites, and both failures are instructive.** Removing the Loaders
+bullet's *label* left the words *skeleton*, *shimmer* and *spinner* standing in its body, so the
+class still read as answered and the plant proved nothing. And the regex that then removed the
+whole bullet was anchored on *"have had"* where the pack says *"never had"* — a fixture that
+changed nothing, caught by the runner's own "the fixture changed nothing" branch.
+
+### An exclusion that held by accident now holds by declaration (B-026)
+
+`pigeonhole`'s nine `--cat-*-ink` tokens sit outside the palette gate's semantic peer set, and
+**the exclusion is correct**. Re-measured with the gate's own OKLab metric: the closest pair is
+**4.92 ΔE** at full colour (`--cat-notification-ink` / `--cat-team-ink`) and **1.24 ΔE** under
+deuteranopia (`--cat-marketing-ink` / `--cat-notification-ink`), against a hard floor of 10 and
+a dichromacy floor of 8. No arrangement of nine hues clears those. The board's 4.42 was slightly
+off; its 1.24 was exact.
+
+What was wrong is *how* the exclusion held: `STATUS_TOKENS` matches by **name**, so an author
+renaming a category token `--danger-ink` got a red gate for the right reason **by luck**, and one
+adding a tenth hue got nothing at all.
+
+The pack now declares it where the categories are enumerated — with those numbers, and with the
+carrier named: the chip's **label word**, already required in Components, plus every ink clearing
+AA against its own tint. The hue is recognition, never identification, and a chip rendered as a
+bare swatch is a defect in this pack.
+
+`validate_excluded_sets_are_declared()` holds the enumeration equal to the shipped tokens **in
+both directions** — a tenth hue with no row fails, a row with no token fails — and requires the
+carrier to be stated **in the section that enumerates the set**.
+
+**That last scoping is the fix its own plant forced.** Searching the whole file let a chip rule
+three sections away satisfy a declaration about the hues: the plant broke two of the phrase's
+three occurrences and the third kept the check quiet. A declaration belongs where the reader of
+that set is standing.
+
+B-017's widening, which this row asked to wait for, shipped earlier the same day. Floor 3835 →
+3839.
+
+### A focus ring was guarded by a mechanism that cannot guard (B-027)
+
+The board asked for 42 hand-restated channel literals to be migrated to relative colour, with
+the rule *"the literal must ship first as a fallback"*. Re-measured, the count is **71 across
+14 packs**, not 42 — it grew with the library and nothing was watching it. And **56 of the 71
+are custom property definitions**, which is what settles the decision:
+
+**For a custom property, a literal above a relative declaration is not a fallback.**
+`--x: rgb(from …)` **parses everywhere** — a custom property accepts almost any token sequence
+— so it wins the cascade unconditionally, and the invalidity surfaces only where the property
+is *substituted*, as invalid-at-computed-value-time. By then the literal declaration is gone
+and every `var(--x)` resolves to `unset`, not to the line above it.
+
+**Twenty-three properties across `ledger` and `showroom` shipped exactly that pattern**, and
+`showroom`'s was **`--ring-focus`**. Its own comment named the stake — *"a dropped declaration
+is an invisible focus indicator, which is the one degradation this library may not ship"* — and
+then relied on the mechanism that cannot deliver it. A twenty-fourth, `ora`'s `--accent-hover`,
+had **no fallback at all**.
+
+All twenty-four are inside `@supports (color: rgb(from red r g b))` now, which is where a
+literal fallback actually survives. `validate_relative_colour_is_guarded()` refuses any
+relative-colour custom property outside such a block, planted with `showroom`'s exact shipped
+pattern. `ora`'s literal is per-theme on purpose: `:root[data-theme="light"]` sets its own
+value, so a single literal carries one theme and needs to.
+
+**The remaining 47 are not migrated, and that is the answer rather than a deferral.** Relative
+colour in a definition needs an `@supports` block per pack, which is machinery in the file a
+reader is told to copy verbatim. The cheaper fix is for each literal to state its source —
+`/* = --on-ink at 72% */`, which `ledger`'s eighteen now do — and for a check to recompute the
+declared relationship, so a token that moves fails in CI instead of drifting silently (B-116).
+
+Floors 3839 → 3863 and 2220 → 2346; the palette count moved because the `@supports` blocks add
+a theme map the palette gate now reads.
+
+### The kit migration was finished, and a guard was complete by accident (B-029)
+
+The board said *15 remain*. Fifteen kits carry no container query — and **eleven of those carry
+no width query at all**, so there is nothing in them to migrate. The actionable set was four
+kits, six blocks, and every one of them is already resolved:
+
+- **`datasheet`** — two blocks, marked *"SELF for the guides, PAGE for the heading"* with the
+  reason for each half.
+- **`field-notes`** — two blocks, marked PAGE and SELF.
+- **`maquette`** — one, SELF, with the reason that wrapping every model block to let it query
+  itself is worse than the query.
+- **`pigeonhole`** — one, and it is a `:root` token switch, which the guard exempts **by
+  design**: `:root` is inside nobody's container, so it cannot be replaced by one.
+
+**Zero unmarked, unexempt viewport queries remain across 29 kits.**
+
+What *was* worth fixing is the guard's reach. It read `*/src/styles.css` and nothing else.
+Measured today: zero width queries live outside `styles.css`, and zero CSS files fall outside
+that glob — so its coverage was complete **by accident**. A `@media (max-width: …)` inside a
+`.tsx` template literal, or a second stylesheet beside the first, would have been invisible.
+
+It now reads every `.css`, `.tsx`, `.ts`, `.jsx` and `.js` a kit ships, refuses a corpus of
+fewer than two files rather than passing on an empty one, and is planted with a width query in
+a `.tsx` template literal. Floor 3863 → 3865.
+
+### Three of five typographic findings did not hold, and one counted comments (B-030)
+
+The row listed five primitives the library specifies by hand. Re-measured today, **three of
+the five are wrong or have closed themselves**:
+
+| The row said | Measured 2026-08-20 |
+|---|---|
+| 0 declare a measure in `ch` | **11** files use `ch`; 12 declare a px measure |
+| fluid spacing 0 of 16, type fluid in 5 | **6** and **15** |
+| `font-display` in 21 files, `size-adjust` in none — *"a guaranteed reflow"* | `font-display` in **three** places, all the **same comment line**; **zero** `@font-face` blocks |
+
+The third is the one worth dwelling on. That count matched **string occurrences** — a token
+named `--font-display: aeonik`, and a comment in `manpage`'s token layer describing its
+reference. There is no `@font-face` anywhere in this library and no `font-display` declaration,
+so `size-adjust` and `ascent-override` have nowhere to attach: **font loading is the consumer's
+layer, and that is a decision rather than a gap.**
+
+`text-box-trim` / `text-box-edge` and `@property` do stand, both at zero.
+
+The skeleton's `## Type` section now carries all three answers with today's numbers and a
+support line each: `text-box-trim` as a **recommendation** rather than a rule because Firefox
+lacks it — and with the reason it matters, that a 60px display at `line-height: 1` is 60px of
+line box and not of cap height; `ch` as the rule a px measure stops obeying the moment the face
+changes; and font loading as explicitly not this layer, so a pack that wants metric-matched
+fallbacks says so in Gotchas instead of shipping an inert descriptor.
+
+`validate_font_loading_stays_out()` holds that last decision — and **strips comments first**,
+because a comment is exactly what made the original count wrong. Planted both ways.
+
+`@property` is filed rather than done (B-117): the token layers are files a reader is told to
+copy verbatim, and a block per token would roughly double them. The narrow version worth pricing
+is the tokens that are actually animated, and that list has to be measured from the kits'
+transitions rather than guessed.
+
+Floor 3865 → 3981.
+
+### Two deferred breakpoints, and neither had the container answer the board assumed (B-032)
+
+Both blocks carried `TODO-CONTAINER B-032` and both were held out of 1.23.0 on the right
+principle: **the breakpoint has to come from the component's own geometry**, and carrying 768
+or 640 into `@container` would be the same mistake in newer syntax. Doing that arithmetic is
+what closed the row — and it did not produce container queries.
+
+**`datasheet`'s instrument grid needed no query at all.** It is already
+`repeat(auto-fit, minmax(180px, 1fr))`: two columns need 180 + 180 + 1px of border = **361px**
+of grid width, and below that `auto-fit` drops to one on its own. The
+`@media (max-width: 640px)` was not helping it — it was **overriding** it, forcing a single
+column on a 640px phone whose instrument may hold two comfortably. Resolved by deleting the
+block.
+
+**`blueprint`'s two derive to numbers that give the game away.** The registration tick is
+`--tick-len: 8px` sitting `--tick-gap: 6px` off each corner, so its two top arms consume 16px
+of the element's width, and at the field's own `--grid-step` of 32px the marks stop reading as
+corners only below **~48px**. The column rule is `--rule-w: 1px` at each edge of a column whose
+inner padding is `--space-8` = 32px — and **this pack states no measure anywhere**, so the only
+derivable floor is the **64px** of padding itself.
+
+48px and 64px are far below anything a container reaches. That is the tell: the 768px
+breakpoint was never about fitting. A registration mark and an edge rule are **ornament**, and
+how much ornament a page carries is a page decision — so both are marked PAGE, with the
+arithmetic written at the block so the next reader does not redo it.
+
+**TODO-CONTAINER is at zero and pinned there.** An escape with no ceiling becomes the habit it
+was meant to interrupt; the ratchet is planted with a marker added back. Floor 3981 → 3983.
+
+### Gate
+
+- **Coverage is pinned, in both directions.** `check_ratio_coverage()` reads
+  `validate_palette.py:ratios` from `test/floors.json`: `computed` may only rise,
+  `unresolved` and `unguarded` may only fall. A check-count floor cannot see this —
+  deleting one `on \u0060--bg\u0060` from a table header moved 27 claims out of the
+  arithmetic while every other check stayed put. Watched failing at 348 → 342.
+- **Five plants, one per bucket**, so the *classification* is watched discriminating and
+  not only the arithmetic: a table declaring a base its numbers contradict, the same
+  claim with the base removed, prose naming no partner, a ratio placed at a gradient
+  stop, and a named partner whose subject cannot be paired.
+- `PARTNER_PHRASE` now reads past a determiner — *"text on the `--accent` fill"* named a
+  partner the gate could not see. The regex moved rather than the sentence.
+- `ARGUED` lists only terms that can reach it. `RATIO_SKIP` already drops a floor, a
+  bound and a rejected candidate one step earlier, so repeating those would have been
+  alternatives that never match — which reads as coverage and is not.
+- Floor `validate_palette.py` 1924 → 1951, all 27 from the newly-checked table claims,
+  then 1951 → 2087 for the declared-set sweep.
+- **Four plants for the declared sets, and the third is the point**: a file-wide
+  *"never by colour alone"* must NOT excuse a pair below the hard floor. Without it the
+  check is a phrase-detector and one sentence silences it.
+- A figure with no qualifier is skipped rather than guessed, and the count of
+  disclosure figures no check could bind prints on every run (7 today, B-055/B-056).
+- The `field-notes` kit's token block was re-copied from the pack's layer, which the
+  copy-never-transcribe check caught on the first run after the comment changed.
+- **`validate_composited_fill_contrast` composites in sRGB gamma space**, because that is
+  where the browser does it. A first draft composited this module's linear values and
+  reported 1.88:1 where the page renders 3.16 — wrong in the direction that fails a
+  correct pack.
+- **Composites are read per CELL.** A Components row is resting | hover | active |
+  disabled, and reading the whole row made `cyclorama`'s hover fill the resting chip's
+  text colour: 1.26:1 reported against a chip that renders at 12.38:1. There is a plant
+  for it.
+- Two live subjects, one of each verdict — which is what makes it a check. Measured
+  first: 44 lines in the library pair a token with a percentage and only two state a
+  composite fill AND a type size. The other 42 are filed (B-058), not guessed at.
+- Floors `validate_palette.py` 2087 → 2092, `computed_at_least` 348 → 349.
+- **Five plants for the prescribed pairs**, including the hover-swap path B-039 is made of
+  and the `as above` inheritance. The fixture had to be corrected before they meant
+  anything: it used B-039's own values, where the resting pair already fails, so three of
+  the five were passing for the wrong reason.
+- A claim added in this change had to be rewritten because `on` and its token landed on
+  either side of a line break — the same refusal B-039's own fix met, and correct both
+  times.
+- Floors 2092 → 2119 and `computed_at_least` 349 → 353.
+- **`scripts/sync-kit-tokens.py`, and `npm run sync-kits`.** A kit's `styles.css` is its
+  pack's token layer verbatim plus its own component half, so every edit to a token
+  layer — down to a comment — makes the kit red until the block is copied again. That
+  copy was done by hand three times today and the third touched **15 kits at once**. The
+  script takes the block boundary from the file's *committed* version instead of guessing
+  it, and a kit that was hand-edited inside its token block reports `STUCK` and is left
+  byte-identical. Watched refusing in a worktree, with the component half intact.
+- Floors 2119 → 2214, `computed_at_least` 353 → 448, `unguarded_at_most` 150 → 57.
+
+### Fixed
+
+- **1.45.0's release notes were stranded under `[Unreleased]`.** The release commit wrote
+  a summary above the accumulated section and left it in place, so the shipped version's
+  detail was labelled unreleased and the next entry would have landed on top of it. The
+  detail is folded into 1.45.0 below, where it belongs. `validate_release_register()`
+  now refuses an `[Unreleased]` heading that is not the first section, and refuses a
+  second one — the same class as the duplicate `## [1.35.0]` found on 2026-08-20, one
+  heading over, and nothing had been looking at it.
+
 ## [1.45.0] - 2026-08-20
 
 **Four packs contradicted their own token layers, and the layer is what an implementer
@@ -36,7 +1202,7 @@ Registers that had stopped:
   at sixteen packs and asserted all 71 had been hand-verified. Measured today: **513 claims,
   336 guarded, 177 unguarded at 22 packs** — and the run prints that pair every time, with
   the hand-verification dated for what it was. Two more restated counts went the same way.
-- **`package.json` said twenty-nine packs and named 27** — `awning` and `bulletin` missing
+- **`package.json` said thirty-one packs and named 27** — `awning` and `bulletin` missing
   from the surface an npm reader chooses from. It is an enumeration site now.
 - **Two `## [1.35.0]` sections** under one version, and four releases in the CHANGELOG with
   no tag. The duplicate is resolved; the untagged four are **reported, not created** — a tag
@@ -63,7 +1229,7 @@ that failed.
 Checks: `validate.py` 3398 to **3543**, `validate_palette.py` 1912 to **1924**, `sloplint.py`
 635. Ten new plants, all caught; 49 planted defects across the three self-tests, 0 missed.
 
-## [Unreleased]
+
 
 ### Four packs contradicted themselves, and six sweeps now say so
 
@@ -164,7 +1330,7 @@ measured, and never gates content.
   the comment says plainly which of them was ever verified.
 - `test/validate.py` said "the six packs shipped before the widening" (seven) and
   "seven of the ten widened packs" (twenty-two widened). The split is printed now.
-- `package.json` said "twenty-nine locked style packs" over a list of **27** —
+- `package.json` said "thirty-one locked style packs" over a list of **27** —
   `awning` and `bulletin` were missing. `ENUMERATION_SITES` gained `package.json`,
   which is the description npmjs.com renders: its count was already policed and
   the names beside it were not.
@@ -288,7 +1454,7 @@ with their numbers.
   Rendered and confirmed at 1440×1000, including the press: the primary's offset goes
   3px → 1px as the control travels 2px, while the secondary beside it keeps its 3px.
 - **Three stale counts in `SURFACE_COMPOSITION.md` that no gate could see.** It said the
-  token names were not uniform "across the twenty-one" when twenty-nine packs ship, and
+  token names were not uniform "across the twenty-one" when thirty-one packs ship, and
   that the accent is `--accent` "in eighteen" when it is twenty-seven — `--brand` in
   `field-notes` and `--cta` in `orchard` are still the only two exceptions, counted. The
   third was a count of packs carrying an `@role non-text:` colour, which is now stated as