fix: the wizards' seed input never arrived (RD-39)

All three wizard containers read `this.seed()` in the constructor. Angular
binds component inputs after the constructor runs, so the value was always the
`initial` default, `seeded !== initial` was always false, and every mount took
the `draftSync.resume()` branch. The `seed` input was dead code.

The two single-step forms built on the same idiom read the input inside the
microtask and work correctly. That contrast is the diagnosis.

Impact: 21 seeded wizard stories rendered step 1 instead of the state they
asked for. Storybook is this repo's UI test surface, so the states with no
other coverage were exactly the ones not rendering — Submitting, Submitted,
Failed, Ingediend, Mislukt. The a11y runner checks that whatever rendered is
accessible, never that the right thing rendered, so nothing caught it.
Production was unaffected: no route binds `seed`.

Read the input inside the microtask, matching the two forms. Turn the spec's
old `componentInstance.dispatch(...)` workaround into a real regression test
through `componentRef.setInput('seed', ...)`.

Verified: with the intake fix reverted the two spec cases fail; with it, 319
pass. `npm run ci --full` is green, and the newly rendered markup produced no
axe violations. A browser check of seven seeded stories across all three
wizards asserts text only reachable from a seed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/project/readable-codebase/RD-39-seed-input-timing.md

@@ -0,0 +1,82 @@
+# RD-39 — The wizards' `seed` input never arrived
+
+Status: done
+Source: found while planning the machine-wiring migration (ADR-0007 arc)
+
+## Why
+
+All three wizard containers read their `seed` input **in the constructor**:
+
+```ts
+const seeded = this.seed(); // always the `initial` default
+queueMicrotask(() =>
+  seeded !== initial ? this.dispatch({ tag: 'Seed', state: seeded }) : this.draftSync.resume(),
+);
+```
+
+Angular binds component inputs **after** the constructor runs. So `seeded` was
+always the `initial` default, `seeded !== initial` was always false, and every
+mount took the `draftSync.resume()` branch. The `seed` input was dead code.
+
+The two single-step forms built on the same idiom prove the diagnosis by
+contrast. They read the input **inside** the microtask, and they work:
+
+```
+change-request-form.component.ts:191  queueMicrotask(() => this.dispatch({ tag: 'Seed', state: this.seed() }));
+besluit-form.component.ts:176         queueMicrotask(() => this.dispatch({ tag: 'Seed', state: this.seed() }));
+```
+
+## Impact
+
+Every seeded wizard story rendered step 1 instead of the state it asked for —
+21 stories across the three wizards. Storybook is this repo's UI test surface
+(CLAUDE.md "Testing"), so the states that had no other coverage were exactly
+the ones silently not rendering: `Submitting`, `Submitted`, `Failed`,
+`Ingediend`, `Mislukt`.
+
+Nothing caught it. `.storybook-ssp/test-runner.ts` runs axe only — it checks
+that whatever rendered is accessible, never that the right thing rendered.
+`intake-wizard.component.spec.ts` had worked around it by calling
+`componentInstance.dispatch(...)` instead of setting the input, which is the
+shape of a test written against a broken input path.
+
+**Production was unaffected**: no route binds `seed`, so the resume branch was
+always the correct one there.
+
+## Decisions
+
+1. **Read `seed()` inside the microtask**, matching the two forms. Six lines
+   across three files. The `if/else` replaces the ternary because the branches
+   are statements, not values.
+2. **Keep the microtask.** It is what defers the dispatch past input binding.
+   The larger fix — `start(seed)` on an application store called from
+   `ngOnInit` — belongs to the ADR-0007 migration, not here. This ticket makes
+   the existing seam correct; it does not move it.
+3. **Turn the workaround into the regression test.**
+   `intake-wizard.component.spec.ts` now mounts through
+   `componentRef.setInput('seed', …)` via a `mountSeeded` helper, and a new
+   `honours the seed input` case asserts the machine state directly. Both fail
+   on the old code, which is the point.
+4. **Do not touch `draftSync.enabled`.** `enabled: () => this.seed() ===
+initial` reads the input lazily inside a lambda called from an effect, so it
+   was already correct. Changing the input to `| null` is part of ADR-0007.
+
+## Verification performed
+
+- `npx ng test ssp` with the intake fix reverted: **2 failed** (both new spec
+  cases). With the fix: **319 passed**. That differential is the proof.
+- `npm run ci --full` green, including 112 storybook a11y tests. The newly
+  rendered markup (`<app-document-upload>`, error alerts, `<app-confirmation>`)
+  produced **no** axe violations — the predicted a11y fallout did not happen.
+- Browser check against the built Storybook, seven seeded stories across all
+  three wizards, each asserted to contain text only reachable from its seed
+  (`Netwerkfout` for `Failed`, `referentienummer` for `Ingediend`, `Documenten`
+  for step 3), plus a negative control that step 1 does **not** show step 3's
+  title. All seven pass.
+
+## Trap for the next person
+
+`Failed`, `Submitted`, `Ingediend` and `Mislukt` are unreachable without a
+network submit, so they can only be produced by a seed. If a future change
+breaks the input path again, those four stories silently fall back to step 1
+and axe still passes. The spec added here is the guard; keep it.

docs/project/readable-codebase/README.md

@@ -133,6 +133,7 @@ two. Note that RD-15 exists because 22 abandoned agent worktrees are still on di
 | RD-36 | `ui/dashboard/` → `ui/overzicht-secties/` + 2 stale `dashboard.page` paths   | 04         | yes       | todo   |
 | RD-37 | **a11y:** 5 suppressions name a closed ticket — decide the `li[…]` host      | 01         | yes       | todo   |
 | RD-38 | One member order for the 3 wizard containers + 2 pure extractions            | 22, 23     |           | done   |
+| RD-39 | **Bug:** the wizards' `seed` input never arrived (21 stories)                | 38         | yes       | done   |
 
 The ID order already respects every dependency, so it is the recommended running order.