fix(cibg): address fieldset padding + atomic-design audit cleanup

Fix: address-fields shipped a scoped `fieldset { padding:0; margin:0 }` that
tied on specificity with CIBG's `.form-horizontal fieldset` and, injected later,
won — flattening the grey-box padding on "Mijn gegevens" and the registratie
wizard address step. Removed the override (kept the legend rule).

Audit cleanup (minimise atoms, converge or document near-duplicates):
- delete dead `app-card` (+ story; zero consumers, superseded by app-data-block)
  and the unused `.app-info-box` utility
- review-section raw <h2> -> app-heading; registratie-wizard lone <dl> -> app-data-block
- keep application-link's raw <h3> (vendored `.applications li a h3` chain needs
  the bare element) with an inline justification
- document the deliberate "don't merge these" verdicts in a new Convergence
  decisions table in atomic-design.mdx; record the card removal in cibg-gaps.mdx

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/docs/atomic-design.mdx

@@ -99,3 +99,27 @@ A new page should be **composition of existing blocks**. Adding a new building b
 exception, not the reflex — if you reach for one, that is a signal to check whether an
 existing atom/molecule already covers it. Fewer primitives → less to test, less to learn,
 one place to fix a bug.
+
+## Convergence decisions — pairs that look duplicated but stay separate
+
+Periodically we audit for near-duplicate blocks. Some collapse into one; a few **look**
+similar but earn their separation. This table records the "don't merge these" verdicts so
+the next person doesn't spend an afternoon re-deciding. (Deliberate CIBG-specific deviations
+live in [CIBG gaps](?path=/docs/foundations-cibg-gaps--docs); the FE⇄DS "same shape, different
+context" cases in [Layers](?path=/docs/foundations-layers--docs).)
+
+| Pair | Why kept separate |
+| --- | --- |
+| `choice-link` vs `application-link` | Share the same `to`/`clickable`/`activate` navigation triad, but bind **different vendored patterns** — CIBG _Keuzelijst_ (`.keuzelijst__link`, `.stretched-link`) vs _Aanvragen_ (`.dashboard-block.applications li a`) — with different list/host semantics (`app-choice-link` renders an inner `<li>`; `application-link` **is** the `<li>`). Merging would fight the vendored CSS. Extract the shared triad into a mixin only if it grows. |
+| `text-input` / `radio-group` / `checkbox` | Share only the standard Angular **ControlValueAccessor** boilerplate (the `writeValue`/`registerOn*`/`setDisabledState` block). They render genuinely different controls, so they stay three atoms. A base CVA class is the only DRY move — a refactor, not a component merge, and not worth it at three. |
+| `button variant="subtle"` (`.btn-link`) vs `app-link` | A subtle button _looks_ like a link but is an **action** (`<button>`, emits click); `app-link` is **navigation** (`<a routerLink>`). Different semantics and a11y roles → keep both. |
+| `shell` / `page-shell` / `wizard-shell` | Three distinct jobs that **compose**, not overlap: persistent app chrome (mounted once) → routed page body → the wizard form/step frame. |
+| Raw `<h3>` in `application-link` vs the `heading` atom | The vendored `.applications li a h3` chain styles the **bare `<h3>`**; wrapping it in the `app-heading` host element would sit between the anchor and the h3 and can break that selector. This is the one sanctioned raw-heading; everywhere else uses `<app-heading [level]>`. |
+
+Single-consumer shared blocks (e.g. `placeholder-chip`, `rich-text-editor`, `checkbox`, the
+`task-list`/`choice-list`/`choice-link` family) currently have one consumer each. They stay in
+`shared` as design-system primitives; relocate one into its consuming context only if it stays
+single-consumer long-term. That is a watch-item, not a merge.
+
+The last audit also **removed** a genuinely dead block — a generic white `app-card` with zero
+consumers (superseded by the grey `app-data-block` as the single data surface).