diff --git a/docs/studies/README.md b/docs/studies/README.md
index 76fb305..6a01d21 100644
--- a/docs/studies/README.md
+++ b/docs/studies/README.md
@@ -92,6 +92,9 @@ language authority.
   feedback from exercising that topology.
 - [Referential example data and read-only preview provenance](referential-example-data-and-read-only-provenance.md)
   records one preview-model experiment.
+- [Carousel and the existing pager](carousel-and-pager.md) collects the
+  accessibility and scroll-snap prior art for the pager/carousel widget
+  candidate and proposes candidate answers to its open ownership questions.
 
 These entries are an inventory, not a roadmap. The index remains as a stable
 router even when its contents change. Removing a leaf means only that the
diff --git a/docs/studies/carousel-and-pager.md b/docs/studies/carousel-and-pager.md
new file mode 100644
index 0000000..3a06ec2
--- /dev/null
+++ b/docs/studies/carousel-and-pager.md
@@ -0,0 +1,313 @@
+# Carousel and the existing pager
+
+- **Status:** Non-normative candidate study for the future `ui` sub-specification
+- **Lifetime:** Disposable study
+- **Method:** Primary-source reading of the referenced specifications retrieved
+  July 20, 2026, plus direct verification against the code both before and
+  after the 0.4 replacement landed on `main` (#24, merge `896aa6b`)
+- **Origin:** The "Carousel and existing pager" candidate in the widget
+  candidate landscape ([#19](https://github.com/gridaco/uhura/issues/19))
+- **Doctrine:** [Mission and identity](../doctrine/mission.md),
+  [authoring ergonomics](../doctrine/authoring.md), and
+  [design principles](../doctrine/principles.md)
+- **Related study:** [Machine-first language, opt-in UI, and explicit framework
+  features](machine-first-language-and-opt-in-ui.md)
+- **Authority:** Research only; this document reserves no syntax, accepts no
+  contract, and creates no obligation for the `ui` sub-specification
+
+## 1. Problem and non-goals
+
+Uhura ships a `pager` element whose observable contract has never been
+settled. The candidate landscape asks five questions about it: the smallest
+renderer-neutral contract; who owns the current page; how keys, inactive
+slides, direction, page change, and static preview work; whether carousel is a
+Pattern over pager, an opt-in compound widget, or a separate Element; and how
+real controls, interactive indicators, autoplay, focus/hover stopping, reduced
+motion, and non-drag alternatives are guaranteed.
+
+This study collects the external prior art, records what the implementations
+actually do today, applies the machine/UI state-ownership boundary to the
+contested state, and proposes candidate answers.
+
+Non-goals:
+
+- deciding `ui` grammar or attribute spelling;
+- specifying autoplay scheduling or animation implementation;
+- settling the Pattern/compound/Element taxonomy itself — this study only
+  places carousel within whichever tiers the `ui` sub-specification defines;
+- touching the deterministic core language.
+
+## 2. What exists today
+
+### 2.1 The v0 spike contract
+
+The 0.3 spike catalog (`examples/instagram/client/catalog/base.toml`,
+removed from `main` when #24 landed and preserved in git history) declares `pager` as a layout viewport with keyed children, an
+`indicator` token (`none | dots`), a required `label`, and a `page-change`
+event of kind `observe`. The catalog is self-describing about its own gaps:
+the current page is "uncontrolled in the spike", and `page-change` is
+"Declared for controlled use; the spike never binds it".
+
+### 2.2 Uhura 0.4 (merged as `main` by #24)
+
+On `main` after the replacement:
+
+- The checked vocabulary keeps `pager` as a native element with attributes
+  restricted to `indicator` (`none | dots`) and a required `label`, a
+  `page-change` event, and **no current-page, page-size, or direction
+  attribute** (`crates/uhura-check/src/ui_catalog/elements.rs` at the current
+  head; earlier revisions carried the same rules inside
+  `crates/uhura-check/src/checker.rs`).
+- The web renderer projects `pager` to `role="group"` plus an `aria-label`
+  taken from `label`, builds a scroll-snap track
+  (`scroll-snap-type: x mandatory`; each child `flex: 0 0 100%;
+  scroll-snap-align: center`), and derives the active page purely inside the
+  renderer as `round(track.scrollLeft / track.clientWidth)`
+  (`web/src/renderer/primitives/pager.ts` after the primitive-adapter
+  refactor). Dots are non-interactive and marked `aria-hidden`.
+- Since the primitive-adapter refactor that landed with #24, the renderer
+  **does** dispatch
+  `page-change`: the pager primitive fires it whenever its rounded active
+  index changes during scrolling. The catalog types the payload as `Unit`
+  (`crates/uhura-check/src/ui_catalog/elements.rs`), so the machine is told
+  *that* the page changed but not *which* page is current. Earlier states of
+  both v0 and 0.4 declared the event while binding it in zero renderers; no
+  example binds it yet.
+
+So the notification half of the contract has just been realized, but the
+position half remains renderer-owned: with a `Unit` payload there is still no
+way for a program to know or control which page is current, and dispatch
+tracks rounded-index crossings during the gesture rather than resting snap
+positions. Which page the machine may own, and in what vocabulary, is exactly
+the question this candidate leaves open.
+
+### 2.3 ARIA APG carousel pattern
+
+The [APG Carousel pattern](https://www.w3.org/WAI/ARIA/apg/patterns/carousel/)
+supplies the accessibility half of the prior art:
+
+- The container takes `role="region"` or `role="group"` with
+  `aria-roledescription="carousel"` and a required accessible name. Each slide
+  takes `role="group"` with `aria-roledescription="slide"` (the tabbed variant
+  uses `tabpanel`), named individually — a "3 of 10" name is explicitly
+  admitted because `group` supports neither `aria-setsize` nor
+  `aria-posinset`.
+- Previous- and next-slide **buttons are listed unconditionally as needed
+  features** (APG's wording is "needed", not "required"). A slide picker is
+  explicitly optional and comes in two shapes: a tabs-pattern picker (single tab stop) or
+  a grouped button picker where the current slide's button is
+  `aria-disabled="true"` (kept in the tab sequence deliberately).
+- A rotation stop/start button is required **only if** the carousel
+  auto-rotates, must be first in the carousel's tab sequence, and carries no
+  toggle state — its label changes with the action instead.
+- Auto-rotation must stop when keyboard focus enters the carousel (and must
+  not resume without an explicit request) and must stop while the pointer
+  hovers the carousel. A live-region wrapper is optional: `aria-live="off"`
+  while auto-rotating, `"polite"` otherwise.
+- Activating rotation/previous/next does not move focus.
+- The pattern says **nothing** about touch or swipe gestures and nothing about
+  reduced motion. The gesture half of a carousel is simply outside APG's
+  scope.
+
+### 2.4 CSS Scroll Snap
+
+[CSS Scroll Snap Level 1](https://www.w3.org/TR/css-scroll-snap-1/) supplies
+the paging-mechanics half, and its own boundary is instructive:
+
+- The normative surface is geometric: snap positions are alignments of a
+  child's snap area within the container's snapport; `mandatory` obliges the
+  container to rest at a snap position when scrolling terminates;
+  `proximity` leaves even that to UA discretion; `scroll-snap-stop: always`
+  forbids passing over a snap position mid-gesture.
+- Everything kinetic is deliberately unspecified: "The CSS Scroll Snap Module
+  intentionally does not specify nor mandate any precise animations or physics
+  used to enforce snap positions; this is left up to the user agent."
+- After content changes, the UA must re-snap, and "if the scroll container
+  was snapped before the content change and that same snap position still
+  exists … the scroll container must be re-snapped to that same snap
+  position". The UA therefore already tracks *which element* is current — it
+  just never exposes it: Level 1 defines no event and no script- or
+  accessibility-visible notion of the snapped element. (Snap events are being
+  drafted in Scroll Snap Level 2; that attribution is from general knowledge,
+  not this document's references.)
+- Axis and alignment resolve logically against the snap container's writing
+  mode, so an inline-axis `start` alignment is the right edge under RTL.
+
+The design lesson both sources agree on: specify the *resting states* and the
+*semantic controls*, and leave the *motion between them* to the realization.
+
+## 3. Ownership analysis
+
+Apply the state-ownership boundary articulated for 0.4 (PR #24 discussion):
+durable, program-observable state belongs to the machine; the UI may own
+ephemeral interaction mechanics; observations flow down; semantic intent
+returns through explicit, typed machine inputs; the UI must not silently
+retain or mutate state on which program behavior depends.
+
+For a pager this splits cleanly:
+
+| Concern | Owner | Reasoning |
+| --- | --- | --- |
+| Current page, when anything observes it | Machine | Conditional content, analytics, guards, or persistence make it program-observable by definition |
+| Current page, when nothing observes it | Renderer | Equivalent to the caret/IME allowance for `textfield`: pure interaction mechanics |
+| Scroll physics, momentum, in-flight gesture offset | Renderer | Exactly what CSS Scroll Snap reserves for the UA |
+| Page set, order, and identity (keys) | Machine (via markup projection) | The keyed children are projected data; re-snap identity depends on stable keys |
+| Indicator rendering (dots) | Renderer | Mechanics today, and correct as mechanics — until indicators become interactive, at which point activation is semantic intent and must dispatch a typed input |
+| Autoplay *policy* (whether/interval) | Declaration | An accessibility-governed behavior, not free renderer choice |
+| Autoplay *scheduling* (timers, pause on hover/focus) | Renderer | Kinetics, like snapping physics |
+
+The one genuinely contested cell is the current page, and the textfield
+precedent already resolves it: renderer-owned while unobserved, machine-owned
+(controlled) the moment the program depends on it. The failure mode of leaving
+it uncontrolled while pretending otherwise is precisely the declared but
+never-bound `page-change` we have now.
+
+## 4. A small operational model
+
+A pager, renderer-neutrally, is:
+
+```text
+pages   : ordered list of keyed slides (keys unique, order meaningful)
+current : key into pages (not an index; see re-snap rule)
+axis    : inline | block, resolved logically (RTL flips inline)
+events  : PageChanged(key) — semantic intent, emitted at rest, not per frame
+invariant : current ∈ keys(pages); pages nonempty ⇒ current defined
+```
+
+Consequences worth making explicit:
+
+- **Key, not index.** CSS Scroll Snap's re-snap rule is element-identity
+  based: if the snapped element survives a mutation, the container re-snaps to
+  *it*, not to its old offset. An index-based contract cannot express that; a
+  key-based one gets it for free, and deleting the current slide becomes the
+  same well-defined question as deleting any keyed row.
+- **Rest, not motion.** `PageChanged` fires when a scroll terminates on a new
+  snap position — mirroring `mandatory`'s end-state obligation — never during
+  the gesture. In-flight offset is renderer mechanics and is not observable.
+- **Deterministic replay.** Because the event carries the resting key only,
+  a trace replays identically regardless of gesture physics, which keeps the
+  self-verifying preview property intact.
+
+## 5. Adversarial cases
+
+- **Current slide deleted.** Machine updates `pages`; the invariant forces a
+  new `current`. Candidate rule mirroring the CSS ambiguity clause: nearest
+  following key, else nearest preceding. A controlled pager makes this a
+  checkable machine decision instead of UA-defined drift.
+- **Stale page intent.** User swipes to a slide the machine has meanwhile
+  removed: `PageChanged(key)` arrives with a dead key. This is exactly the
+  0.4 result taxonomy's `stale` outcome; no new machinery is needed.
+- **Single page.** Previous/next controls have nothing to do; the APG
+  expectation that they exist collides with a degenerate page set. Candidate:
+  controls render disabled rather than vanish, keeping the tab order stable.
+- **Autoplay under reduced motion.** APG is silent on reduced motion; CSS
+  leaves motion to the UA. If the platform signals reduced motion, candidate
+  rule: declared autoplay is suppressed and the optional live region behaves
+  as "not automatically rotating" (`polite`). This must be a conformance
+  rule, not renderer goodwill, precisely because neither upstream source
+  covers it.
+- **RTL.** If `axis` resolves logically, "next" under RTL moves left. The
+  contract must say whether previous/next controls are logical (recommended,
+  matching Scroll Snap's writing-mode resolution) or physical.
+- **Mid-gesture machine write.** The machine sets `current` while a drag is in
+  flight. Renderer owns the gesture; candidate rule: the write wins at gesture
+  end unless the gesture itself terminated on a different snap position after
+  the write — the same last-writer question every controlled input has, and
+  it should be answered once, in the `ui` sub-specification, for all
+  controlled mechanics.
+
+## 6. Static checking and runtime consequences
+
+Checkable today, with no new machinery:
+
+- keyed slides with a checked key type (the 0.4 checker's ui `each` key
+  rules already exist);
+- required `label` (already enforced in the 0.4 checker);
+- `page-change` bound ⇒ payload type matches the declared machine input;
+- **the reverse obligation**: an element must not declare an event in the
+  checked vocabulary that no conforming renderer realizes — the v0 pager
+  fails this, and the 0.4 primitive-adapter refactor now satisfies it, which
+  demonstrates the obligation is realistic to enforce.
+
+Checkable once controlled current-page exists:
+
+- controlled/uncontrolled consistency: a pager whose current page is read
+  anywhere must bind `page-change` (the analogue of the controlled
+  `textfield` spelling);
+- autoplay declared ⇒ rotation control present in the realization contract.
+
+Runtime: stale keys map to `stale`, malformed payloads to `invalid`, guarded
+refusals to `blocked` — the existing result vocabulary covers the pager
+without extension.
+
+## 7. Renderer and boundary effects
+
+- **Web realization** keeps exactly what it has: scroll-snap track, physics,
+  in-flight offset, dot synchronization, and the newly wired unit
+  `page-change`. The candidate changes the dispatch to a position-bearing
+  payload emitted at rest, and adds the option of driving position from a
+  controlled `current`.
+- **Static preview pose.** The candidate landscape's framing lists "a
+  scrolled viewport" as a pose problem. A key-based `current` doubles as the
+  declarative pose: a static preview of a pager at slide `k` is well-defined
+  without any gesture machinery. The same shape should generalize to
+  `scroll`'s preview pose.
+- **Accessibility projection.** The APG mapping is mechanical from the model:
+  container name from `label`; per-slide names ("n of m" admitted); controls
+  as real buttons. Today's realization stops at `role="group"` +
+  `aria-label` with no roledescription, no slide names, and no controls —
+  the gap between the two is a candidate conformance checklist, not a design
+  question.
+- **Editor.** The editor canvas needs the pose, the key list, and the
+  indicator declaration — all present in the model; nothing editor-specific
+  leaks into the contract.
+
+## 8. Migration
+
+- v0/0.4 markup (`label`, `indicator`, keyed children) is forward-compatible
+  unchanged; every addition (controlled `current`, bound `page-change`,
+  `axis`) is opt-in.
+- The observe-kind `page-change` of the spike maps onto a typed machine input
+  under the 0.4 event model, following whatever general observe→input mapping
+  the `ui` sub-specification adopts for the other declared events.
+- No existing example binds `page-change`, so nothing breaks by defining it
+  properly; the instagram spike gains the option of making story/media paging
+  observable.
+
+## 9. Candidate answers to the issue's questions
+
+1. **Smallest renderer-neutral contract:** keyed slides + resting current
+   page (key) + logical axis + required label + indicator token. Everything
+   kinetic excluded, mirroring the Scroll Snap normative split.
+2. **Current page ownership:** renderer-owned while unobserved;
+   machine-owned (controlled, with `PageChanged` at rest) the moment any
+   program behavior depends on it — the textfield rule applied to paging.
+3. **Keys, direction, preview:** keys mandatory and identity-bearing
+   (re-snap follows the element, so must the contract); direction logical
+   with RTL resolved as in CSS writing modes; static preview = declared
+   current key, which also answers the pose problem.
+4. **Carousel's tier:** carousel adds only *semantic controls and rotation
+   policy* on top of pager mechanics — prev/next buttons, optional picker,
+   rotation control — all of which are ordinary buttons plus one declaration.
+   The evidence therefore favors **composition over a new Element**: a
+   Pattern (or compound) over `pager` + `button`, in whichever tier the `ui`
+   sub-specification provides for named compositions. A separate Element is
+   justified only if autoplay policy cannot be declared compositionally.
+5. **Guarantees:** APG's baseline features (real prev/next buttons; rotation
+   control iff autoplay; focus/hover stopping; no focus theft) become
+   conformance obligations of the carousel composition. Reduced motion and
+   touch alternatives are **not covered by either upstream source** and are
+   exactly where Uhura's contract must go beyond prior art: candidate rules —
+   reduced motion suppresses declared autoplay; every gesture-reachable page
+   is control-reachable (the required buttons already guarantee this).
+
+## 10. Open questions
+
+- The general observe→typed-input mapping for `ui` events (this study assumes
+  it; the sub-specification owns it).
+- The last-writer rule for machine writes during in-flight gestures, shared
+  with all controlled mechanics.
+- Whether the composition tier ("Pattern") the answer to question 4 relies on
+  is itself part of the `ui` sub-specification's initial scope.
+- Whether `indicator` grows an interactive form (picker), which would move
+  indicator activation from mechanics to semantic intent.