# Changelog

## 0.1.15 — 2026-09-18

**Two real defects in `.admin-table`, found building a full product against this
checkout** (a Beakon 2.0 reskin — six registers growing to ~90 pages). Documented
first as findings (`docs/beakon-reskin-findings`, still open); this is the actual
code fix for the two that are genuine `src/` bugs rather than page-local
workarounds.

### Fixed

- **Row hover never showed on an even row in dark mode, at any colour.**
  `.admin-table tbody tr:nth-child(even)` picks up a `body[data-theme="dark"]`
  prefix for its own dark value, reaching specificity (0,3,3) — one type-selector
  heavier than the base `.admin-table tbody tr:hover` (0,2,2) can ever reach, dark
  mode or not. So in dark mode specifically, hover always lost on an even row
  regardless of what colour a consumer set it to; verified in a browser before the
  fix, an even row's background never changed on `:hover` at all. A new
  `body[data-theme="dark"] .admin-table tbody tr:hover` rule ties the zebra rule's
  own (0,3,3) and wins by coming later in the same sheet — the same specificity
  technique already used for the nth-child/selected-row collision, applied here
  against hover instead of selection. It reads `--px-hover-btn` (10% white in
  dark), the token every other interactive surface in this theme already hovers
  to, rather than inventing a new dark value for whatever colour hover used in
  light.

- **`--praxis-color-blue-10` had no dark remap at all**, unlike every sibling tint
  rung in the palette (teal, green, orange, red all get one). It stayed `#edf0fd`
  — near-white — in both themes. Neither of its two real consumers in `src/` was
  actually exposed by this (`--praxis-tone-info-bg` and `.admin-banner` both
  already restate their own dark value directly, bypassing the alias), so this is
  a consistency fix for the token itself — closing the one gap in an otherwise-
  complete set — rather than a fix to a known-broken shipped component. Given
  `#131c33`, following the same recipe already used for teal-10: a dark, muted
  rung of blue-10's own hue, not the brightened accent hue `--praxis-color-blue-60`
  uses for dark-mode accents.

  Both verified against a rebuilt `dist/praxis.css` with a headless-browser
  `getComputedStyle` probe, in both themes: the dark fix changes only the dark
  hover on an even row (odd-row hover, light mode entirely, and both
  `--praxis-tone-info-bg`/`.admin-banner` are all unchanged before and after).

Six fixes surfaced building a real product on `.admin-table`, `.tbtn`, `.cn-item`,
checkboxes, `.tb-dropdown__item` and a toolbar search field — all found the same
way as the dark-mode row-hover and blue-10 fixes below: exercising the shipped
components on real content, not from re-reading the source.

### Fixed

- **A selected `.admin-table` row rendered the zebra stripe instead of its own
  colour, on every even-indexed row.** `.admin-table__row--selected` ships at
  (0,1,0)/(0,2,0); `tbody tr:nth-child(even)` is (0,2,2) in light and (0,3,3) in
  dark — heavier either way, so the stripe always won on an even row regardless of
  sheet order. Fixed with the wrap-ancestor trick already used elsewhere in this
  file: `.admin-table-wrap .admin-table tbody tr.admin-table__row--selected`
  reaches (0,3,2), enough for light; dark's zebra rule is a type-selector heavier
  still, so it needed its own `body[data-theme="dark"]`-prefixed copy at (0,4,3).

- **`--admin-row-selected` (`#fbe6c2`) sat a few units from
  `--praxis-tone-warning-bg` (`--praxis-color-orange-10`, `#fdf2e6`)**, so a
  selected row could swallow a "Medium"/"In progress" warning pill it was meant to
  let you keep reading. Every other selected/active state in the system is teal —
  `.cn-item.is-sel`, `.cf-cond.is-selected`, `.admin-tab--active` — so
  `--admin-row-selected` now aliases `--praxis-color-teal-10` instead of carrying
  its own near-duplicate colour, and resolves correctly in both themes with one
  declaration.

- **A primary button rendered as `.tbtn.tbtn--primary` painted a near-white wash
  under white label text on hover.** The generic tool-hover rule,
  `.tbtn:not(.tbtn--run):not(:disabled):hover`, excluded `.tbtn--run` and
  `:disabled` but not `.tbtn--primary`, `.btn--primary` or `.pill-btn` — and at
  (0,5,1) it beat the primary's own `:is(...):hover` rule's (0,3,1) outright,
  since `:is()` takes only its highest-weight argument. Added the three missing
  exclusions.

- **`.tbtn` and `.cn-item` kept their hover underline when rendered as `<a href>`,
  unlike every other anchor-capable component.** `.ws-item:hover` and
  `.praxis-navrail__link:hover` both set `text-decoration:none`; these two never
  did, because their own pages only ever render them as `<button>`. Closed on
  both, matching the rest of the system.

- **Checking a table-row checkbox could resize the row.** The checkbox is
  `display:inline-grid` with no `vertical-align` set, so it sits on the text
  baseline — and an empty `inline-grid` synthesises that baseline differently from
  one with a real child (the `:checked` tick, via `::after`). Measured 46px
  unchecked vs. 45px checked in a bare single-row table. Fixed with
  `vertical-align:middle`, which does not conflict with the toggle switch's own
  `:not(.switch):not(.switch *)` exclusion.

- **`.tb-dropdown__item:hover` and a toolbar search field both under-delivered
  their intended surface treatment.** `--px-hover`'s 2.5% wash is tuned for a bare
  row on `--px-surface`; a dropdown item is already on its own tinted menu
  surface, so the same 2.5% read as almost no change. Reached for
  `--praxis-color-teal-20` instead — one step past the `teal-10` selection fill,
  so hover and selected stay distinct (`--praxis-color-blue-10` was considered
  first, but it's the system's *info* tone and a hovered item would have read as
  a callout). Separately, `.filter-toolbar-row .filter-search-input input` sits
  on `--px-field` (`#EEF1F4`, a step off white meant to read recessed one level
  inside a `--px-surface` card) directly on `--px-surface-2`/`--px-page` — three
  close enough values that the field nearly disappeared. Given the surface it was
  missing plus `--px-tool-shadow`, matching how a floating toolbar control
  already lifts off the band under it.

## 0.1.14 — 2026-09-01

**Two things the ≤768px search panel got wrong.** Both were only ever visible on a
phone or a tablet, and both were in the panel 0.1.12 introduced.

### Fixed

- **A hairline ran down the middle of the joined search field.** The term input and
  its magnifier button are meant to read as one control — matching radii, no gap.
  Each carried its own `box-shadow:0 0 0 1px` ring, and a ring is drawn around the
  whole of every element that has one, so the two of them painted a 1px line
  straight down the seam they were supposed to hide.

  Both halves now take a `border` instead and drop the side they share
  (`border-right:0` on the input, `border-left:0` on the button), so the pair has a
  single continuous edge and nothing down the middle. The two shared corners were
  already square, so no radius is lost. Dark mode moves with it: row 1 keeps its
  `--px-edge` ring, the joined pair takes `--px-edge` as a border colour.

  `box-sizing:border-box` is now declared on both rather than assumed —
  `praxis-reset.css` is optional for consumers, and without it a 1px border would
  have made these 46px tall in a 44px row.

- **Opening the panel buried the page under a suggestion list.** `.ss-pop` opens on
  the term input's `focus`, and opening the panel focuses that input — so the one
  gesture that reveals the search field also fired the "Recent searches" dropdown
  over the content, every time. Above 768px the pill is always in the bar and
  focusing it is a separate, deliberate act; in the panel it is not.

  `.appbar__search .ss-pop{display:none}` inside the ≤768px block. Suppressed
  rather than re-timed: there is no gesture left at this width that means "show me
  suggestions" and not "show me the field". The rule is scoped to the bar's own
  popup, so a page using `.ss-pop` elsewhere is untouched, and the drivers only
  toggle `[hidden]` — never an inline `display` — so no `!important` is needed.

## 0.1.13 — 2026-08-31

### Fixed

- **A Lucide icon could be repainted into a solid blob by a consumer's own
  `.icon` rule.** Lucide writes `fill="none"` as a *presentation attribute*, and
  a presentation attribute loses to any CSS rule at all. `praxis-admin.css` has
  carried `.icon{width:22px;height:22px;fill:currentColor}` since it was written
  — correct for the inline sprite SVGs it was written for, which are fill-drawn,
  and silently wrong for every stroke-drawn Lucide icon that also carries
  `.icon`. On the admin pages that filled fifteen icons, the app bar's magnifier
  among them, so the same search field read as a solid glyph there and an
  outlined one everywhere else.

  `praxis-lucide.js` now injects `svg.lucide{fill:none}` alongside the
  stroke-width rule it already ships. `svg.lucide` is specificity 0,1,1 against
  `.icon`'s 0,1,0, so it wins outright and does not depend on sheet order — and
  because it travels with the converter that generates these SVGs, every
  consumer is covered, not just the admin sheet. Verified against the built
  package on three groom-lake pages: filled Lucide icons went 15 → 0 on the
  admin pages and 1 → 0 on the search page, with stroke-width unchanged at 1.5.

  **If you were relying on filling a Lucide icon**, you no longer can via
  `.icon`; set `fill` on a more specific selector than `svg.lucide`.

## 0.1.12 — 2026-08-28

**The app-bar module selector survives narrow widths.** Below 768px the search
pill carried two controls with room for one, and the sheet chose the term input:
`body:not([data-page="search"]) .msel{display:none}`. So on a phone you could
search, but not choose what you were searching — and the scope stayed on whatever
it had been, still filtering, with nothing on screen saying so.

Neither control stays in the bar now. The whole pill collapses behind a magnifier
and reopens as a panel under the bar: module selector on the first row, term input
on the second.

### Added

- **`praxis-appbar-search.js`** — owns the toggle and
  `body[data-appbar-search="open"]`. No markup contract: it injects the magnifier
  into `.appbar__right`, so any page already shipping `.appbar__search` gets the
  behaviour by loading the script. Author `[data-appbar-search-toggle]` in the bar
  to place the button yourself and the script adopts it. Closes on Escape
  (returning focus to the toggle), on click-away, and on growing back past the
  breakpoint. The module selector's body-level menu and the shared `.px-menu` /
  `.select-menu` surfaces are excluded from click-away, since picking a module
  would otherwise dismiss the panel.
- **`.appbar__searchtoggle`** in `praxis-appbar.css` — the magnifier, plus the
  ≤768px panel geometry. The panel splits into two rows by wrapping with a 100%
  flex-basis on the selector, not by a wrapper element, so whatever a host puts
  beside the input lands on the second row with it. Both selector forms are
  handled — `.msel` and `.appbar__module`.

### Fixed

- **A date set from a Standard filter row never reached the expression.** The
  expression-tree date button carries both `.filter-date` and
  `data-role="cf-value-date"`, and there is a handler bound to each. `bind()`
  attaches one document listener per selector, and `stopPropagation()` does not
  stop other listeners on the same node — so both ran, legacy first. It created
  the `input[type=date]` and attached its `change` listener; the tree handler
  then found an input already present, took its `if (!input)` branch as false,
  and never attached the listener that writes to the tree. The value went to the
  flat `filterState`, producing an applied chip with nothing in the readout.
  Dates set in the Custom builder were unaffected — those buttons sit outside
  any `[data-filter-name]`, so the legacy handler bailed on its own guard.
- **The filter modal's resize handles did nothing.** `initModalResize()` bound
  correctly and wrote `--filter-modal-h` on every `pointermove`; nothing read it.
  A grep for the property across the package returned exactly one hit — the
  write. It was orphaned by the stable-height change, which replaced the panel's
  `max-height` with an explicit `height` and hardcoded `min(720px, …)` while its
  own comment claimed `initModalResize()` "already writes that same property".
  The rule now reads `var(--filter-modal-h, 720px)`, so the resting size is
  unchanged and the drag works.
- **`Status` could never appear in the expression.** It carried
  `control: 'segmented'`, which made `isScopeFilter()` true, and scope filters
  live on the flat `filterState` by design. It is an ordinary rich list field
  now — operator + multi-select, like `Priority`. **Its modal row changes from
  segmented Open / Closed / Cancelled buttons to a condition cluster.**
- **An unset value looked like a chosen one.** A condition with no value yet
  rendered as a plain `.cfx-val` — teal semibold, reading as a value literally
  named "…". It emits `.cfx-val--pending` and takes the `.cfx-incomplete`
  treatment.

### Changed

- **`formatExpr()` emits one element per part, and one element per value.** The
  field name was a bare `<b>`; a multi-select was a single `.cfx-val` reading
  `[Aberdeen, Bergen, Calgary]` with brackets and commas baked into the string.
  Neither could be given a shape. The name is now `.cfx-field` and each value is
  its own `.cfx-val`. **Anything selecting `.filter-expr b` or parsing that
  bracketed string will need updating.**
- **Spacing between the readout's parts is `margin-inline` on the part classes**,
  not literal spaces in the markup — a space glyph is whatever width the font
  makes it, and it collapses unpredictably once a part becomes an inline-flex
  box. The default 2px a side reproduces the previous gap; a consumer painting
  the parts as chips raises it.
- **A quick-filter card shows 7 values, and "See More" opens a popup.** It was 6,
  and "See More" expanded the rest in place — a card with 30 values grew the card
  and with it the whole strip, pushing results off screen. It now opens
  `openMultiSelectMenu`, the same primitive the drawer's value dropdowns use, so
  the list arrives searchable with the same checkbox rows and dismiss behaviour,
  positioned at body level (which matters: `.quick-filters` is `overflow:hidden`
  for its collapse transition and would have clipped an in-card panel). Toggling
  routes through `toggleFilterValue`, exactly as clicking a card row does, so the
  two surfaces cannot drift. `data-action="toggle-quick-more"` is now
  `data-action="quick-more"`; the "See Less" state is gone.
- **`openMultiSelectMenu` takes an optional `opts.label`** for per-option display
  text, so the popup can carry the card's `value (count)` labels while still
  handing raw values back on toggle. A no-op for every existing caller.
- **The nav rail's selected item wears the gradient + elevation its neighbours
  already had**, in pink. It was a flat `pink-50` fill — the one button in the
  rail with neither, which read as unfinished beside the others.
- **`.msel` is no longer hidden below 768px**, and `.appbar__right .iconbtn-ghost`
  is no longer hidden on the Search page. Both rules existed to buy horizontal
  room for a pill that is no longer in the bar, so the secondary icon buttons come
  back on phones. **If you relied on either being hidden at that width, they are
  not any more.**
- **`praxis-appbar.css` releases `praxis-module-selector.css`'s narrow-width
  clamp** on `.msel` (`max-width:132px`, 96px at ≤480px) inside the panel. That
  clamp sized the selector to share a narrow bar; it owns a full row here. The
  clamp still applies to a `.msel` used outside `.appbar__search`.

### Housekeeping

- **The repository moved to `IdeagenLtd/shared-praxis`** in the 2026-08-27 org
  migration. `Ideagen-AX/praxis` no longer exists. The package name is unchanged
  — it is still `@ideagen-ax/praxis`, and only the source repo moved — but
  `repository`, `homepage` and `bugs` in `package.json` shipped the old URL in
  every tarball up to 0.1.11, so anyone following a link from npm landed
  nowhere. Fixed here, along with the links the reference site renders.
- **npm's trusted publisher had to be re-pointed by hand.** The OIDC
  `job_workflow_ref` claim names the repository, so publishing from the new home
  fails an authentication check until the trusted publisher on npmjs.com lists
  `IdeagenLtd` / `shared-praxis`. Nothing in `publish.yml` names the org — the
  claim comes from wherever the workflow runs — so this is a one-time change on
  npm's side, now written into that file's header for the next reader.
- **Every GitHub Actions `uses:` is pinned to a full-length commit SHA**, with
  the release in a trailing comment. The IdeagenLtd org rejects a workflow that
  references a moving tag, and the run dies at "Prepare all required actions"
  before a single step executes — so `ci.yml`, `pages.yml` and `publish.yml` were
  all inert on the new remote, publishing included. Pinned at `checkout@v7.0.1`,
  `setup-node@v7.0.0`, `setup-python@v7.0.0`, `configure-pages@v6.0.0`,
  `upload-pages-artifact@v5.0.0`, `deploy-pages@v5.0.0`.
- **`setup-node`'s `registry-url` is gone from `publish.yml`, which is what was
  actually breaking the release.** It writes an `.npmrc` carrying
  `//registry.npmjs.org/:_authToken=${NODE_AUTH_TOKEN}`, and a trusted-publishing
  job has no `NODE_AUTH_TOKEN` by design. npm found an auth line, considered the
  registry credentialled, sent an empty token rather than performing the OIDC
  exchange, and the registry answered the anonymous `PUT` with a 404 naming the
  package. The registry is not lost with it removed: `npm publish` defaults to
  registry.npmjs.org and `publishConfig.access` still carries `public`.
- **`publish.yml` fails on a missing OIDC token instead of on its symptom.**
  Without one, npm does not stop — it publishes anonymously and the registry
  answers `E404 ... PUT /@ideagen-ax%2fpraxis`, which reads as "no such package"
  and sends the reader off to check the name, the scope and the trusted
  publisher, none of which is wrong. A preflight step now asserts
  `ACTIONS_ID_TOKEN_REQUEST_URL` is present and names the actual cause. The
  `id-token: write` grant is also restated at job level.
- **The page template no longer hardcodes the repo URL.** It sat beside
  `build-site.py`'s `GITHUB` constant as a second source of truth, which is how
  two links survived the move pointing at the old org. The template reads
  `{{github}}` now.

## 0.1.11 — 2026-08-26

**The reference site is restructured onto the EHSQ-E design system's information
architecture.** Four content sections — **Foundations, Patterns, Components,
Resources** — each with an overview, taken from
`ehsqe-design-system-docs.vercel.app` so that a reader who knows that site can
find things here. 66 pages became 82.

**Components are grouped by layer, not by stability.** The nav used to read
`Components — ready` / `— settling` / `— unstable` / `— planned`, which answered
"how much will this churn" before "what is it" and moved a page in the sidebar
when its churn risk changed. Stability is now the pill on the page and nothing
else; `group: global | core | records` places the page. Layers are alphabetical,
because a layer is a catalog of thirty things you arrive at knowing the name of;
foundations, patterns and resources keep `order:`, because they are a reading
sequence.

No `src/` behaviour changed. Two live defects that the restructure's new gates
caught are fixed below.

### Fixed

- **The toast interactive demo never worked.** It loaded
  `<script src="../praxis-toast.js">` from inside its example frame, which
  resolves to nothing — so every button on it called an undefined
  `praxisToast()`. `praxis-toast.js` is now a known `data-scripts` value and the
  path is generated.
- **The breadcrumb demo's home link pointed at a page that does not exist.**

### Added — Patterns

- **Ten pattern pages**, a section Praxis did not have: Forms, Navigation, Page
  layout, Data display, Feedback, Loading states, Empty states, Error handling,
  Record page, plus an overview. Each states the **Problem** before the
  **Solution**, which is the EHSQ-E order and the useful part — it lets a reader
  tell whether they have the problem before reading the answer.
- **A five-section skeleton for them, gated:** Problem, Solution, Accessibility,
  Guidelines, Related components. A pattern answers a different question from a
  component; forcing it into the twelve would produce an "Anatomy" of a
  composition.
- Loading states, empty states and error handling **moved from Components to
  Patterns** and were restructured onto that skeleton. They read badly as
  components because none of them is one thing — "loading" is a skeleton, a
  spinner and an inline resolution, and choosing between them is the whole design
  decision.

### Added — Resources

- **Getting started** (moved from Foundations), **For AI agents**,
  **Contributing**, **Glossary** and **Changelog**. The changelog page is
  generated from this file, so it cannot disagree with it. Helix alignment was
  deliberately not added back: per-component Helix sections were removed on
  2026-08-25 and re-adding a comparison the project decided to drop would be a
  regression.

### Added — build gates

- **No internal link may be dead**, and **no internal link may go through a
  redirect stub**. Renaming nine pages left eighteen dead hrefs, every one of
  which still rendered a page that looked fine until it was clicked. A stub is
  for a URL someone else has already shared, not for this site's own navigation.
- **Nothing in the built output may resolve to a 404.** A separate crawl over
  every emitted file, examples included, because the prose check skips
  `<template>` markup — which is exactly where the toast bug was hiding.
  `<code>` and `<pre>` are excluded: prose about a broken `src=` is not itself
  broken, and the check flagged the sentence describing the bug it had just
  found.
- **Pattern pages must follow the pattern skeleton.**
- **`group:` must be one of the three layers.** A typo would quietly create a nav
  group of one.

### Changed — Foundations

- Split to match the reference IA: `space-radius` became **Spacing**,
  **Borders** and **Grid and layout**; `elevation-motion` became **Elevation**
  and **Motion**. Each gained the Usage guidelines and Accessibility sections
  that convention ends on.
- Renamed: `mental-model` → **Design principles**, `type` → **Typography**,
  `icons` → **Iconography**. All ten EHSQ-E foundation names now exist.
- Praxis's eight additional pages — Theming, Materials and glass, Forced colors,
  Print, Right-to-left, Gaps, Corrections, Conventions — stay, after the ten.

### Changed — site

- Ten `redirect_from` stubs keep every old URL alive.
- Section indexes are exempt from `tier:`, `group:`, the skeleton and their own
  breadcrumb, and are excluded from the catalogs so they cannot list themselves.
- The agent doc's planned-page exclusion now keys on `tier`, not on the section.
  It keyed on the section until this release; a planned page now lives under
  `Core` with everything else, so a section test would have started shipping
  thirty phantom classes into the npm tarball.

## 0.1.10 — 2026-08-26

**Every token declaration now lives on `:root`.** Praxis declared 127 of them on
`<body>` — the 41-property `body[data-variant="praxis"]` block and three dark
blocks — and nine of those were *second* declarations of tokens
`praxis-tokens.css` had already made on `:root`. The variant selector outranks
`:root`, so for those nine the token file stated a value that never rendered:
`--praxis-radius-card` read `20px` and rendered `12px`.

That was not only untidy. A custom property is substituted at the element where
the **declaration** lives, so a token declared on `<body>` is invisible to any
`:root` alias of it: `:root{--a:var(--b)}` plus `body{--b:x}` computes `--a` on
`:root` against the old `--b`, and `<body>` inherits that. Five tokens shipped
broken this way. Keeping every declaration on one element makes the class
impossible rather than merely absent.

Overriding at a variant scope is a reasonable pattern when there is more than one
variant. There is not — the frozen "Miramar" comparison view was pruned on
2026-08-12 and `data-variant` has been required-and-always-`praxis` since. The
layer was buying nothing and costing the ambiguity.

**No markup change, and no consumer migration.** The theme attribute is still on
`<body>`. The dark blocks moved to `:root:has(body[data-theme="dark"])`, which
lets `:root` see an attribute it does not carry; the ~220 component rules keyed on
`body[data-theme="dark"] .foo` are untouched, because they only *read* tokens.

### Fixed

- **`--praxis-color-status-info` rendered `#4766eb` in light — a blue the palette
  no longer contained.** It aliases `blue-60` on `:root`, and the variant block
  redefined `blue-60` from `#4766eb` to `#4361c4`, so the alias froze at the
  pre-variant value. Only the dark half had ever been restated (0.1.5, 2026-08-18),
  which left the light half broken for another eight days. It is now `#4361c4`,
  matching `blue-60` as the token file always claimed.

  **This is the only rendered value that changes in this release.** Verified in
  headless Chrome, `getComputedStyle` on `<body>` across 220 tokens x 2 themes,
  before and after: one change in light, zero in dark.

### Changed — `src/`

- The nine variant overrides are folded into `:root` in `praxis-tokens.css`:
  `--praxis-color-text-primary`, `text-secondary`, `border-default`, `blue-60`,
  `--praxis-radius-card`, `--praxis-elevation-1`, `-card`, `-card-raised` and
  `-popover`. Five redundant restatements that already agreed with the token file
  were dropped.
- **The 27-token `--px-*` material layer moved into `praxis-tokens.css`.** It was
  the single most common reason someone grepped the token file for `--px-surface`
  and concluded it did not exist. The dark half stays in `praxis-core.css`, also
  on `:root`.
- `--praxis-color-purple-60` is a real palette rung rather than a variant-only
  extra.
- `body[data-variant="praxis"]` no longer declares a single token. It holds the
  dot-grid page material and nothing else.
- `praxis-workspace.css` and `praxis-admin.css` each carried a body-scoped token
  block too; both moved to `:root`. The workspace one is a drifted duplicate of
  the base dark remap in `praxis-core.css` and two of its values disagree with
  core — preserved verbatim, because this release is a pure move, and flagged in
  place for a separate fix.

### Added — build gates

- **`praxis_meta.body_declared_tokens()` — no token may be declared on `<body>`.**
  The structural invariant behind all of this. Four responsive layout hooks set
  inside `@media` are exempt and named explicitly.
- **`praxis_meta.variant_overrides()` must stay empty** — no token declared twice.
- **`frozen_aliases()` was wrong in two ways, both of which made it report clean
  while a real defect was live.** It only looked at *dark* remaps, so it could not
  see the variant axis at all; and it treated "the token is restated on `body`
  somewhere" as a clean bill of health, which is exactly how `status-info` stayed
  half-broken after being half-fixed. It now works from declaration *sites* across
  every sheet, and a restatement only excuses a rung declaration whose conditions
  it is a subset of. Against the 0.1.9 tree the new version reports the one real
  freeze; the old version reported none.

### Site

- The colour page's *One declaration site* section replaces *The Praxis variant
  overrides nine tokens*, with a redirect-free anchor change and every
  cross-reference updated. The theming, materials, space-and-radius, card and
  card-and-page pages no longer describe a layer that is gone.

### Also in this release — the reference site

**The reference site is now built in Praxis.** It used to consume the tokens and
draw its own chrome from a parallel `--doc-*` layer with hand-written dark
values; it now loads `dist/praxis.css` and wears the real shell — `.app`,
`.appbar`, `.adminnav`, `.content`, `.pageheader`, `.toolbar`, `.admin-body`,
`.admin-card`, `.admin-table`, `.admin-pill`, `.tbtn`, `.px-skip`,
`.px-navtoggle`, and Material Symbols ligatures through `praxis-lucide.js`. Only
prose typography, the table of contents, the example frame and the colour
instruments are still the site's own. No `src/` behaviour changed for that; the
one fix below is what building it that way turned up.

The document sits on its own sheet — `.doc` is an `.admin-card` on the raised
tier — so anything on it steps to `--px-surface-2` rather than trying to lift
off white with a shadow. That is the step Praxis already uses for
`.admin-panel` inside an `.admin-card`, and the token moves away from the sheet
in whichever direction is visible in each theme.

### Added — site only

- **One twelve-section skeleton on every component page, enforced by the build.**
  Anatomy, Variants, States, Responsive behavior, Interactive demo, Code, Markup
  contract, Token reference, Figma adaptation, Usage guidelines, Accessibility,
  Dimensions — from the previous EHSQ-E design system docs, with *API* replaced by
  *Markup contract* since Praxis has no props or events to document, and *Helix
  alignment* dropped. All 18 pre-existing component pages were restructured to it
  with their prose preserved and demoted to `<h3>`, and the five sections none of
  them had were written from `src/`.

  That pass recorded five live defects it found, listed in `ROADMAP.md` — the one
  worth repeating here is that **`.admin-tab` ships with no `:focus-visible`
  rule**, so a keyboard user gets the UA outline over a 2px bottom border.

- **All 35 roadmap items now have a page** — 32 component stubs under a
  *Components — planned* nav section, and three foundation pages for print,
  forced-colors and an RTL cost audit.

- **A `planned` component tier, and the first eight stub pages.** `ROADMAP.md`
  inventories Praxis against six design systems and lists 35 patterns it does not
  define; the eight Tier 1 entries now have pages. They sit in their own
  **Components — planned** nav section and follow the thirteen-section convention
  the previous EHSQ-E design system docs used, with *API* replaced by *Markup
  contract* because Praxis ships no framework bindings.

  These are **excluded from `PRAXIS-FOR-AGENTS.md`**, which ships and is contracted
  to state what Praxis defines rather than what it intends.

### Fixed

- **`.admin-pill--ok` and `.admin-pill--lock` failed WCAG 1.4.3.** The pill label
  is 12px at weight 700, which is not large text (large starts at 18.66px bold),
  so it needs 4.5:1. Both variants were built from palette rungs rather than the
  audited `--praxis-tone-*` pairs: `--ok` was green-60 on green-10 at **3.81:1**
  and `--lock` was red-60 on red-10 at **3.88:1**. Both now take the tone pair,
  one rung darker on the ink over the same tint — **4.99:1** and **5.08:1**.
  `--off` was already 4.89:1 and is left where it is rather than moved for
  symmetry.

  The dark overrides are kept rather than deleted even though they now restate
  what the tone tokens resolve to: `praxis-core.css` remaps the tone pairs under
  `body[data-variant="praxis"][data-theme="dark"]`, so on a dark page that is
  *not* the Praxis variant the pill would otherwise keep its light tint.

### Added

- **`.admin-pill--info` and `.admin-pill--warning`.** The pill had green, neutral
  and red and no blue or amber. The reference site's own admin example was
  already reaching for `.admin-pill--warning` and getting the bare shape with no
  colour, which is how this surfaced. Both take the audited tone pairs, so they
  need no dark override.

## 0.1.9 — 2026-08-21

**Four things the prototype had four copies of.** A toolbar menu, a panel icon
button, a filter field and a toast — each one duplicated across consumer pages,
in versions that had drifted apart. Nothing here changes an existing Praxis
component; it adds the four that were missing, which is why they kept being
written locally.

### Added

- **`.tb-menu` / `.tb-dropdown` — the toolbar menu** (new sheet,
  `praxis-controls.css`). A
  `.tbtn` that opens a panel under itself: Group, sort and column menus, row
  actions, a layout picker. Twelve pages in the prototype carried a copy in
  three incompatible versions — 12px radius and no border in one, 36px items and
  a hairline in another, `__divider` here and `__sep` there.

  This one closes a real gap rather than merely deduplicating: both
  `praxis-toolbar-compact.css` and `praxis-toolbar-compact.js` already keyed off
  `.tb-menu`, so Praxis was reading a class it never defined. Surface follows
  `.px-pop` exactly, so a menu opened from the toolbar and one opened from the
  nav rail are the same object. Both separator names are kept as an alias —
  cheaper than renaming a hairline in seven files.

  Members: `__item`, `__item--danger`, `__sub`, `__sec`, `__divider` / `__sep`,
  `__empty`, and `--right` for a trigger at the end of the band. Styling only:
  opening, closing, focus and Escape are the consumer's.

- **`.iconbtn` — the 34px panel icon button** (`praxis-controls.css`). Expand all,
  pin, a filter toggle, a collapse chevron: the square that sits in a card
  header rather than in the page toolbar band. Not `.tbtn--icon`, which is 40px
  and belongs beside a labelled toolbar button — the two have never been the
  same size on any screen carrying both. `.iconbtn--on` is the pressed state, in
  the app's pink selection accent.

- **`.filterfield` — type-to-filter beside a list** (`praxis-controls.css`).
  Trailing glyph, because the field refines the list under it rather than
  starting a search; the leading-glyph forms are the app bar and `.cn-find`.
  Sized to `.iconbtn` so a header row of both lines up.

- **`praxis-toast.js` — the transient confirmation.** `praxisToast('Saved')`,
  with `tone` and `duration` options, stacking, and a `role="status"` live
  region created at load rather than with the first message — creating a region
  and its text in the same frame is the reliable way to have the announcement
  dropped, which is what three of the four copies did. Styles itself, on the
  `praxis-lucide.js` argument: a component with no markup until it fires cannot
  be allowed to half-install.

- **`praxis-controls.css`** is the new sheet the three CSS components live in.
  `praxis-admin.css` is where `.tbtn` is and would have been the obvious home,
  but it is the largest sheet in the system and carries the application shell,
  a bare-element reset and its own `.tbtn` — eight of the twelve pages carrying
  a local `.tb-dropdown` do not load it, and adding it to pick up a dropdown
  would restyle their whole toolbar. Consumers of the `praxis.css` barrel get
  it either way.

### Changed

- **`.iconbtn` and `.filterfield` are 8px, not the copies' 9px.**
  `--praxis-radius-sm`. A one-pixel difference, and a design system should not
  ship a value off its own scale to preserve it.

- Consumers adopting `.tb-dropdown` in place of a local copy get **36px minimum
  item height** and a focus ring. The four-page group that used 9px vertical
  padding and no ring will see rows a few pixels taller.

### Docs

Two new pages — **Toolbar menu** and **Toast** — and `.iconbtn` / `.filterfield`
join Buttons and Form controls. 31 pages, still at 100% class-family coverage.

## 0.1.8 — 2026-08-20

**Reverts the filter-panel control borders from 0.1.6 and 0.1.7.** A design
decision, taken with the contrast cost understood and accepted.

### Reverted

- **`praxis-filters.css` and `praxis-toolbar-compact.css` return to their 0.1.5
  state**, byte for byte. Nine control boundaries go back under the 1.4.11 3:1
  floor — `neutral-20` at 1.57:1 for `.segmented`, `.filter-type-buttons`,
  `.filter-select`/`.filter-date`, `.filter-search input`, `.filter-view-switch`;
  `neutral-30` at 2.01:1 for `.qfilter__check`, `.select-menu__check`,
  `.cf-group__add button`; and `border-default` at 1.26:1 for
  `.tb-options__search input`.

  **Why.** The app's search pill is drawn with a `rgba(16,36,58,.08)` shadow ring
  rather than a border, so it was never in scope and never moved. Raising the
  filter panel to 3:1 while the pill stayed put made the two read as different
  systems sitting side by side, and the panel is the more visible of the pair.
  Weighed against that, a conformant filter panel that looks broken next to the
  app's primary control was judged the worse outcome.

  The fill cannot rescue this: `--px-surface-2` measures **1.08:1** against
  `--px-surface` in light and 1.12:1 in dark, so the border is the only thing
  identifying these controls and 1.4.11 lands entirely on it. Reaching 3:1 by
  fill alone needs roughly `#767b80`, which is not a light theme. There is no
  version of this that is both conformant and visually unchanged — that is the
  actual trade, and it was made deliberately.

  **The real fix is a design change, not a token change.** Moving these controls
  to a treatment that carries contrast on one edge — a filled field with a 3:1
  underline — would conform without the four-sided weight, but it has to happen
  to the search pill and the filter panel together or the mismatch simply moves.

### Kept

Everything outside the filter panel stands: `--praxis-color-border-strong` at
`neutral-50` (3.84:1), `--praxis-color-status-warning` at `orange-70`, and
`--praxis-tone-warning-fg` at `orange-90`. `border-strong` still serves
`.sm-opt__box`, `.save-field__input` and `--px-check-stroke` in consuming apps,
none of which were part of the objection.

## 0.1.7 — 2026-08-20

**One line of CSS, for a reason worth writing down.** 0.1.6 raised the filter
checkboxes' border from 2.01:1 to 3.84:1 and left the thickness at 2px. Contrast
and thickness multiply, so the visible weight nearly doubled and the controls read
as heavy — reported from review, not from a measurement.

### Fixed

- **`.qfilter__check` and `.select-menu__check` drop 2px → 1.5px.** They were the
  only checkboxes in the system at 2px; `.sm-opt__box`, the 18px box used across
  the record pages, has always been 1.5px. So this is the house convention being
  applied, not a contrast trade-off — the border still measures 3.84:1 on
  `surface-default` and 3.55:1 on `surface-subtle`, unchanged from 0.1.6.

  Worth stating the general rule, because 0.1.6 got it wrong: raising a border's
  contrast to clear 1.4.11 raises its visible weight by the same factor. If the
  stroke was already thicker than it needed to be, the fix lands twice. Check the
  thickness at the same time as the colour.

## 0.1.6 — 2026-08-20

**`src/` changed.** Three colour tokens move and nine control borders darken, all of
it WCAG 1.4.11 (non-text contrast). Light mode only — every dark value is unchanged.

**What a consumer will notice.** Checkbox boxes, segmented controls, select/date
triggers and search inputs get a visibly darker outline; warning icons and warning
state-borders go one step deeper in the orange ramp; warning text on a warning chip
darkens a lot. Nothing moves in dark.

### Fixed

- **`--praxis-color-border-strong` was 2.95:1 on white** — a rounding error short of
  the 3:1 floor for a UI-component boundary. It is the control-boundary token
  (checkbox boxes, input underlines, `--px-check-stroke`), so the floor applies.
  Repointed from `neutral-40` to `neutral-50`: 3.84:1 on `surface-default`, 3.36:1 on
  `surface-subtle`, 4.09:1 on the dark card, so it needs no dark remap of its own.

  Repointed rather than editing the `neutral-40` rung, because
  `--praxis-color-text-disabled` resolves to that same rung and must not move with it.
  It is still `#8898a9`.

- **Nine control boundaries reached past the token to a palette rung**, and every one
  landed under 3:1 — `neutral-20` at 1.57:1 or `neutral-30` at 2.01:1. On an unchecked
  `.qfilter__check` or `.select-menu__check` the border *is* the control, so a 2.01:1
  box is the whole affordance failing. All nine now use `--praxis-color-border-strong`:
  `.segmented`, `.qfilter__check`, `.select-menu__check`, `.filter-type-buttons`,
  `.filter-select` / `.filter-date`, `.filter-search input`, `.filter-view-switch`,
  `.cf-group__add button`, and `.tb-options__search input`.

  `.filter-select:hover` moved to `neutral-60`, because its old `neutral-30` hover was
  now *lighter* than the resting border and the state read backwards.

- **`--praxis-color-status-warning` was 2.69:1 on white.** This token draws icons,
  dots and 2px state borders, all of which owe 3:1, so `orange-60` failed at its own
  job. Now `orange-70` — 3.64:1 on `surface-default`, 3.18:1 on `surface-subtle`, and
  still legibly orange. Dark is untouched at `#ffa32e`.

- **`--praxis-tone-warning-fg` was 3.30:1 on its own chip.** The `tone-*-fg` family is
  the text tier and the other four sit near 5:1 on their backgrounds, but orange at the
  `-70` rung is simply too light to get there. Now `orange-90`, 6.28:1 on
  `--praxis-tone-warning-bg`. That matches the rule an earlier consumer-side audit
  already landed on — use the `-90` step for text on a `-10` tint.

  Note the division of labour: `--praxis-color-status-warning` for the graphical layer,
  `--praxis-tone-warning-fg` for text. Using the status ink as a text colour on a chip
  is what failed, and it still will.

### Added

- **`redirect_from:` in content metadata**, and a stub at
  `/foundations/colour.html` pointing at `/foundations/color.html` — that URL 404'd
  from the 0.1.5 rename. Pages serves static files only, so the stub does what a 301
  would: `meta refresh` for the no-JavaScript case, `location.replace` because it
  fires sooner and adds no history entry so Back behaves, a visible link if both are
  blocked, and `noindex` with a canonical tag so the stub stays out of search results
  while the real page keeps any ranking. The build fails if an old path collides with
  a real page.

## 0.1.5 — 2026-08-18

**`src/` changed, and dark mode changed with it.** Ten fixes: six colour tokens, the
toggle switch, disabled buttons, the Create New catalog's reachability, and a usage
comment that told you to do the wrong thing. Every one was found by the reference site
this release also adds — several of them within minutes of a page first rendering.

**What a consumer will notice.** A page that looks right in light and uses
`--praxis-color-interactive-active`, `border-subtle`, `surface-muted`, `status-info`,
`status-danger` or `interactive-hover` **will render differently in dark**, because
those six were wrong there. Nothing changes in light. Toggle to dark and look before
you upgrade.

### Fixed

- **Four tokens had a dark value that could never apply.** Each was declared on
  `:root` as `var(--rung)` while the rung's dark remap was declared on `body`.
  Custom-property substitution happens at the element where the *declaration* lives,
  so each was computed on `:root` against the light rung and `body` inherited that.

  | Token | Was, in dark | Now |
  |---|---|---|
  | `--praxis-color-interactive-active` | `#135d63` — a light-mode ink | `#5CE0E5` |
  | `--praxis-color-border-subtle` | `#edf0f2` | `#222b39` |
  | `--praxis-color-surface-muted` | `#edf0f2` | `#222b39` |
  | `--praxis-color-status-info` | `#4766eb` | `#7a93e0` |

  `interactive-active` is the one that mattered: in dark, a control's resting state
  was bright cyan `#29D2D7` and its pressed state resolved to a dark `#135d63`, so
  pressing it made it go *darker*. `--praxis-color-surface-subtle` aliases the same
  `neutral-10` and never had the bug, because it was always given its own dark value
  rather than relying on the alias to carry the theme — that is the pattern.

  `praxis_meta.frozen_aliases()` detects this class structurally and is now a build
  gate at zero. It is deliberately not derived from resolved values: a resolver asked
  for the dark value substitutes the dark rung and reports a difference the browser
  never produces, which is why this went unseen.

- **`--praxis-color-interactive-hover` restated for dark** (`#42D9DE`), which is a
  deliberate widening of the fix. It was not frozen — it aliases `teal-70`, which has
  no dark treatment — but fixing `active` alone would have produced a triad that goes
  bright, then dark, then bright. The progression now mirrors light in *direction*:
  each interaction step increases contrast against the ink on the fill. Measured with
  `--px-primary-fg`: light `#fff` 4.50 → 5.88 → 7.57; dark `#08313a` 7.48 → 8.08 →
  8.77. `#42D9DE` is the one invented value, the midpoint of the two it sits between.

- **`--praxis-color-status-danger` failed contrast on dark.** At `#e22d38` it measured
  **3.50:1** on the dark card against the 4.5:1 WCAG 1.4.3 asks of text, while
  success, warning and info sat at 7.88, 7.89 and 5.30 — it was the only status ink
  with no dark value. Now `red-40` (`#ed7b82`), an existing rung, at 5.81:1.

- **The toggle switch had two markup forms and only one was styled.**
  `praxis-admin.css` defines `.switch` as a wrapper containing `.track` and `.thumb`;
  `praxis-filters.css` reads `.switch:checked`, which only matches with the class on
  the input — and `praxis-filters.js` emits that second form plus `.switch__track`
  and `.switch__thumb`, **neither of which any stylesheet defined**. `class="switch"`
  on the input also excludes it from the Praxis checkbox rules, so the filter
  drawer's On/Off row rendered as a bare browser checkbox between two labels.

  Both forms are now defined from one set of values and render identically —
  verified at 34×20 with `neutral-30` unchecked and `teal-60` checked in all four
  permutations. The **wrapper form is canonical**, because `praxis-core.css` is
  already committed to it: its checkbox exclusion is
  `:not(.switch):not(.switch *)`, and that second clause only means anything if
  `.switch` can have descendants. The sibling form was kept working rather than
  removed because `praxis-filters.css` and its script agree with each other, and
  rewriting either would mean touching JavaScript to fix something CSS could close.
  `.switch__track` / `.switch__thumb` are the canonical part names; `.track` /
  `.thumb` still work and are kept for consumers, but they are about as
  collision-prone as class names get in a shared sheet.

- **Disabled buttons had no appearance.** `praxis-core.css` excludes `:disabled` and
  `[aria-disabled="true"]` from the hover wash and the press transform, so a disabled
  button stopped responding while looking identical to an enabled one at rest.
  `.tbtn` and `.admin-ghostbtn` now take `opacity:.45`, `cursor:not-allowed` and no
  shadow on both. Opacity rather than a flat grey, because it has to work on
  `.tbtn--primary`'s gradient. The other eight named button classes still have no
  base at all, so their disabled state remains yours.

- **`praxis-create-new.js` was unreachable under a bundler.** It declared
  `CREATE_CATALOG` and `CN_TEMPLATES` with top-level `const` and never assigned them,
  which in a classic script creates a global *lexical* binding — readable as a bare
  identifier from another classic script, but not a property of `window`, so
  `if (window.CREATE_CATALOG)` always failed. And because `package.json` declares
  `"type": "module"` with no `export` anywhere in the package, importing the file
  through a bundler made both arrays module-scoped and invisible, with no error:
  `import '@ideagen-ax/praxis/dist/praxis-create-new.js'` was a silent no-op. Both are
  now on `window`; the bare identifiers still work.

- **`praxis-dotfield.js` rendered nothing if you followed its own usage comment,**
  which showed `create` → `setMode` → `setParam` and never mentioned `start()`, which
  the loop requires. The comment now includes it, plus `restart()`, `destroy()` and
  the fact that the dots are drawn white and need a dark backdrop.

### Added

Nothing below here changed CSS or JS. It is the reference site and the machinery
behind it — which is what found every fix above.

- **A reference website, at <https://ideagen-ax.github.io/praxis/>.** Every token
  viewable in both themes, and a page per component with live examples. Deployed
  from `main` only by `.github/workflows/pages.yml`; branches and pull requests get
  no public URL, only the `build-site.py --check` gate in CI.

  Two properties make it trustworthy rather than decorative. It renders the real
  `dist/`, built from `src/` at deploy time, so there is no second copy of Praxis
  to keep in step. And each example's live frame and source panel come from the
  same `<template>`, so they cannot disagree — the failure mode of every
  hand-maintained gallery.

  Resolved token values are read in the browser off two hidden probe documents,
  one per theme, rather than computed by the build. A build can only report what a
  token is *declared* as, and this is a system where that is regularly the wrong
  answer.

- **`build-site.py`** — the generator, plus `--serve` (rebuild on each request),
  `--check` (the CI gate), `--coverage` (undocumented class families) and
  `--agents-doc` (render the content to markdown). Stdlib-only, like the other two
  build scripts.

- **`site/content/`** — the prose, one HTML file per page. This is now the source
  for component documentation. `PRAXIS-FOR-AGENTS.md` stays hand-written and keeps
  shipping in the tarball until the site covers every component, at which point it
  becomes generated from the same content.

- **`praxis_meta.py`** — one measurement implementation, shared by `build-ds.py`
  and `build-site.py`, so the site and `DESIGN-SYSTEM.md` cannot state different
  numbers. It also gained cycle detection over `var()` chains, and extraction of
  the `--px-*` material layer and the nine `--praxis-*` tokens that
  `praxis-core.css` overrides under the Praxis variant.

- **Color blocks that draw rather than tabulate.** The full palette as one grid
  (hues down, rungs across, light over dark in each cell, so which rungs the dark
  theme remaps is visible at a glance); each hue as a continuous ramp with its rung
  and hex on every step and the label ink flipping where the rung crosses over; and
  the semantic layer drawn by role — inks as text, borders as a hairline, fills as a
  control — in both themes, each sample on the surface its own theme provides.

- **Every remaining component page — the site now documents all 234 class families
  in `src/`.** Twelve new pages: buttons, form controls, nav drawer and rail flyouts,
  card/page/texture, Create New, module selector, quick-filter rail, compact toolbar,
  workspace chrome, breadcrumb back button, the admin shell, and the page-family and
  part-only names in core. Plus two new foundation pages — naming and state
  conventions, and the corrections to `DESIGN-SYSTEM.md` — and depth on filters
  (the custom-filter expression tree) and Mazlan (the drawer inventory).

- **`PRAXIS-FOR-AGENTS.md` is now generated** from `site/content/`, and grew from 918
  hand-written lines to 3,799. It still ships in the tarball; `site:check` fails if the
  committed file is stale.

- **Class-family coverage is a gate, not advisory.** It went from 15% to 100% while
  those pages were written. It also now checks whether a page *claims* a family in its
  `classes:` metadata rather than whether the string appears anywhere in a content file
  — the looser test counted `.card` as documented because seven pages mentioned it in
  prose while it had no page and no example. Counts are distinct rather than per sheet;
  the old report said 197 where 177 families were real.

- **`praxis_meta.frozen_aliases()`**, which finds tokens whose dark value can never
  apply: declared on `:root` as `var(--rung)` while the rung's dark remap lands on
  `body`. Four tokens are in that state. Detected structurally, because asking a
  resolver would report a difference the browser does not produce. Advisory in the
  build output, not a gate, since it is a defect in `src/` rather than in the site.
  **Now a gate**, at zero, since the four it found are fixed in this release.

### Fixed — measurement

- **The class-family measurement counted `url()` paths and quoted strings.** A
  scan for `.name` read `url(fonts/Gilroy-Regular.woff2)` as a class called
  `.woff2` and an SVG namespace as `.w3` and `.org`, on top of the `.css` in prose
  that comment-stripping already had to handle. This inflated the apparent surface
  of the system from 234 real class families to about 370, and it showed in the
  component inventory in `DESIGN-SYSTEM.md`, where `praxis-reset.css` and
  `praxis-tokens.css` appeared to define classes they do not.

- **The filter drawer's On/Off toggle renders as a bare browser checkbox.**
  `praxis-admin.css` defines `.switch` as a wrapper containing `.track` and `.thumb`;
  `praxis-filters.css` reads `.switch:checked`, which only matches with `.switch` on the
  input — and `praxis-filters.js` emits that second form, plus `.switch__track` and
  `.switch__thumb`, neither of which is defined in any stylesheet. `class="switch"` on
  the input also excludes it from the Praxis checkbox rules, so nothing styles it at
  all. Documented on the form-controls page with both forms side by side; the fix
  belongs in `src/`.

- **`praxis-create-new.js` is unreachable under a bundler.** It declares
  `CREATE_CATALOG` and `CN_TEMPLATES` as top-level `const` and never assigns them to
  `window`, so they are global lexical bindings that another classic `<script>` can read
  but `window.CREATE_CATALOG` cannot. `package.json` declares `"type": "module"` and no
  shipped script has an `export`, so importing that file makes both arrays module-scoped
  and invisible, with no error.

- **`praxis-dotfield.js` renders nothing if you follow its own usage comment.** The
  comment shows `create` → `setMode` → `setParam` and never mentions `start()`, which
  the loop requires. The dots are also drawn white, so it needs a dark backdrop.

- **`.mazlan-mark--xl` is documented in `PRAXIS-FOR-AGENTS.md` as 40px and does not
  exist.** It appears only inside a comment in `praxis-mazlan.css` describing one
  consumer's own page, so using it silently gets the 20px base. The live example on
  the Mazlan page is what caught it.

- **`DESIGN-SYSTEM.md` was one `var()` usage stale** — 1,501 against a real 1,502,
  from the 0.1.4 profile-menu fix landing without `npm run docs` being re-run.

## 0.1.4 — 2026-08-14

### Fixed

- **The profile menu scrolls instead of running off the bottom of the window.**
  `.profile-menu__pop` had no height limit, and its contents are set by the page
  rather than by the component: admin pages add six "Switch to" links, the
  record pages add the Appearance row, the workspace adds "Viewing as". On a
  laptop viewport the tallest of those extended past the bottom edge, so Sign
  out — the last row — could not be reached at all. The panel is now capped at
  the room below the app bar, `calc(100vh - var(--praxis-appbar-h) - 24px)`,
  and scrolls its overflow, matching the cap `.cn-flyout` already used.

  It scrolls the panel itself rather than an inner body, as the module selector
  does, because there is no sticky header in this menu to hold in place.
  `overscroll-behavior:contain` keeps a flick past the last row from scrolling
  the page underneath.

## 0.1.3 — 2026-08-13

**No CSS or JS changed.** `src/` is byte-identical to 0.1.2, so nothing renders
differently — `dist/` differs only in its version stamp. This release exists to
put a build guide in the tarball and to make publishing reproducible.

### Added

- **`PRAXIS-FOR-AGENTS.md`, and it now ships in the package.** The doc a
  teammate — or their coding agent — reads to build a prototype: the canonical
  app shell as literal markup, per-component markup contracts, the measured
  token list, and the classes Praxis references but never defines. It was
  missing from `files`, so installing from npm rather than cloning got you the
  README and nothing else.

  Three things it records that were not written down anywhere:
  `praxis-admin.css` owns the app shell every page needs (`.app`, `.main`,
  `.content`, app-bar positioning, the rail container, `.tbtn`, `.switch`,
  `.mazlan-mark`, `box-sizing`) despite its name; `.btn` has **no base
  definition** in the package, only `.btn--primary`; and the Mazlan drawer
  cannot be used at all from the package, because `praxis-mazlan.js` needs
  ~20 fixed element ids whose markup is not shipped and it no-ops silently
  without them.

- **Automated publishing.** A `v*` tag now builds, verifies and publishes via
  npm trusted publishing — OIDC, no token, provenance attached. Previously
  impossible from CI: the npm account uses a passkey, so a non-TTY publish died
  with `EOTP`.

- **CI on every pull request**, running the build's own verification — no font
  binary in the package, no surviving `--ehsq-*` token, no CDN-breaking relative
  path, no unresolvable `var()` in the barrel.

### Fixed

- **The README pinned `0.1.0` while the package was `0.1.2`**, so anyone
  copying the quickstart's CDN URLs got a two-versions-stale build. The CI job
  now warns when the two disagree.

- **`npm run check` was documented as the CI gate** in both the README and
  `CLAUDE.md`. It cannot be: it compares a rebuild against the `dist/` on disk,
  and `dist/` is gitignored, so in any fresh clone it exits 1 with `dist/ is
  missing`. It is a local staleness check. CI runs `npm run build`, which
  performs the same verification.

## 0.1.2 — 2026-08-13

The first release with new material in it. Extracting Praxis made visible how
much of the system the consuming prototype was still carrying itself: 88 custom
properties defined in page `<head>`s. Most were aliases for tokens that already
existed here. These were not.

### Added

- **Motion**: `--praxis-motion-slowest` (420ms), the one step of a competing
  page-level scale that had no equivalent in `--praxis-motion-*`.
- **Easing**: `--praxis-ease-spring-out`, `--praxis-ease-spring-bouncy`. The
  spring family was two-thirds present; pages supplied the rest.
- **Quick-rail motion**: `--praxis-rail-duration`, `--praxis-rail-ease`,
  `--praxis-rail-travel`. Seven pages defined these identically.
- **Menu motion**: `--praxis-menu-duration`, `--praxis-menu-ease`.
- **Glass material**: nine `--praxis-glass-*` tokens, light and dark. A frosted
  translucent surface three pages each had a slightly different recipe for.
- **Status tones**: ten `--praxis-tone-{neutral,info,success,warning,danger}-{bg,fg}`
  pairs, light and dark.
- **Layout**: `--praxis-record-rail-w` (300px) and `--praxis-control-h` (32px),
  each identical everywhere it appeared.

### Fixed

- **The teal scale had no dark treatment.** `--praxis-color-teal-80` is a
  light-mode ink and measured 1.55:1 on a dark panel — illegible — while
  teal-10/20 are near-white and glared as light blocks. Four pages were each
  patching this locally. Now remapped in the dark theme, alongside the status
  inks. **This changes dark rendering anywhere teal-10/20/80 is used**, which is
  the point: those surfaces were displaying the bug.
- **The token-counting regex matched BEM modifiers.** Unanchored, it read
  `.chip--danger:hover{…}` as a token named `--danger`. This overstated every
  published count (206 against 195 real) and, more seriously, let phantom
  definitions mask genuinely undefined tokens in the build's own verification.

### Changed

- `--praxis-duration-fast/base/slow`, added earlier the same day, are **removed**
  before anyone could depend on them. They were a third motion scale introduced
  on a misreading: `--praxis-motion-drawer:280ms` is a single drawer value, not
  evidence of a 180/280/420 scale. The canonical scale is `--praxis-motion-*` at
  120/180/260, already referenced 213 times.

## 0.1.1 — 2026-08-13

Praxis now lives in its own repository, `Ideagen-AX/praxis`, and `src/` is the
source of truth. It was extracted from the groom-lake prototype with history —
174 of that repo's 334 commits touched these files — and the sheets are
byte-identical to the ones 0.1.0 shipped.

**No functional change.** The built output differs from 0.1.0 only in two
provenance strings that name the new source, plus the version-pinned Lucide
fallback URL. Nothing a consumer renders changes. This release exists to prove
the pipeline publishes from the new home before anything depends on it.

Six files stayed with the prototype because they are bound to that application
rather than to the design system; `dist/manifest.json` lists them with reasons.

## 0.1.0 — 2026-08-13

First packaged release. Praxis has existed in the prototype for months; this is
the first time it is consumable from outside it.

### Packaging

- `dist/praxis.css` bundles the foundation and all 13 component sheets in cascade
  order. Individual sheets ship alongside it for selective import, with subpath
  exports for the common entry points.
- `dist/` is generated by `build-package.py` from `prototype/`, which stays the
  single source of truth. Nothing is hand-copied, and the build verifies its own
  output before it can be published.
- The licensed Gilroy `@font-face` blocks are stripped on the way out and replaced
  by `praxis-fonts.example.css`. No font binary ships in this package.
- Lucide is bundled and version-pinned. `praxis-lucide.js` previously fell back to
  `unpkg.com/lucide@latest`, an unpinned dependency that could change under
  consumers without warning.
- `praxis-reset.css` ships but is excluded from the bundle, so Praxis cannot
  restyle a host application's bare elements unless asked to.

### Fixed while packaging

Building the package surfaced defects that were invisible inside the prototype,
because two page `<head>`s happened to supply what the shared sheets were missing.

- **The Mazlan drawer had no transition on any page but one.** Its opacity and
  transform transitions referenced `--dur-base`, `--ease-spring` and
  `--ease-spring-soft` with no fallback, and those three were defined only in
  `contextual-awareness.html`. Everywhere else the declarations were invalid at
  computed-value time, so the drawer snapped instead of sliding. The values are
  now canonical tokens (`--praxis-motion-drawer`, `--praxis-ease-spring`,
  `--praxis-ease-spring-soft`).
- **33 alias-token references across seven shared sheets** pointed at a
  `--t-*` / `--s-*` / `--d-*` / `--dur-*` / `--ease-*` / `--color-*` vocabulary
  defined only in page `<head>`s. On the other 23 pages they silently resolved to
  their hard-coded fallbacks. All now use canonical `--praxis-*` tokens; every
  substitution was checked to equal the fallback it replaced, in both themes.
- **`--appbar-height` was canonicalised to `--praxis-appbar-h`.** Three shared
  sheets read it while only two pages and the legacy sheet defined it.
- `--t-sm`'s fallback was `.875rem`, which is `--praxis-type-size-base`, not
  `--praxis-type-size-sm` (`.8125rem`). Mapping by name rather than by value would
  have quietly changed that text size.

### Known limitations

See *Stability* and *Known rough edges* in the README: the filters sheet themes by
flipping palette primitives, `praxis-admin.css` duplicates chrome geometry
locally, and two host-supplied layout knobs are read but not defined here.

### Publication

Published to npmjs.com as `@ideagen-ax/praxis` under MIT, public access.

The scope is the design team's own (`ideagen-ax`), not the company-wide
`ideagen` — which is already taken on npm — and it matches the existing
`Ideagen-AX` GitHub org. The predecessor packages (`@ideagen-ehsqe/*`) went to
GitHub Packages instead; this is the first Ideagen design-system package on the
public registry.