refactor: one member order for the 3 wizard containers (RD-38)

RD-22 and RD-23 brought the wizard containers under the 250-line budget, so
`max-lines` reports nothing. The files still read badly. Line count was never
the problem.

Fix three things in all three containers:

1. The member order was scrambled, and it differed per file. `registratie`
   declared `draftSync` in the middle of a run of `computed`s; `herregistratie`
   read `this.stepLabels.length` seven lines before `stepLabels` existed; the
   three files put the copy arrays in three different places. All three now use
   one nine-section order, so they compare side by side.
2. Pure logic sat in the container. Extract `digitalDocumentIds` into
   `upload.machine.ts` — the "digital and finished uploading" filter was
   written out four times, and it removes a `documentId!` assertion from both
   containers. Extract `diplomaMsg` into a sibling of the step files.
3. Comments carried archaeology. Drop the three RD-05 references and keep the
   rule. Drop "replaces sessionStorage" and the note about focus management that
   moved to the shell. Fix `intake`'s class comment, which claimed answers
   persist to sessionStorage and was contradicted 30 lines below.

`phase` deliberately stays in all three: it cannot live in `domain/`, and three
siblings plus three specs is a worse trade than 17 readable lines. The store ⇄
`draftSync` cycle also stays — both callbacks are deferred, so it is safe, and
one comment now names it.

No behaviour change. Member lists and every `private`/`protected`/`readonly`
modifier are unchanged, which the showcase depends on.

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

docs/project/readable-codebase/RD-38-wizard-container-readability.md

@@ -0,0 +1,100 @@
+# RD-38 — One member order for the three wizard containers
+
+Status: done
+Source: user report — "the registratie-wizard component still doesn't look readable"
+
+## Why
+
+RD-22 and RD-23 split the intake and registratie wizards into step components. That brought
+every container under the 250-line budget, so `max-lines` reports nothing. The user read the
+result and still called it unreadable. Line count was never the problem.
+
+Three problems remained, and the user named all three:
+
+1. **The member order is scrambled, and it differs per file.** You bounce up and down to
+   follow one thread, and the three containers cannot be compared side by side.
+2. **Pure logic sits in the container**, where it has no test.
+3. **Comments carry archaeology** — ticket numbers, history, and one comment that describes
+   code which is no longer in the file.
+
+RD-02 measures a file. This ticket is about what a file reads like at a fixed size.
+
+## Evidence, before the change
+
+- `registratie` declared `stepLabels`/`stepTitles` between the `seed` input and
+  `state`/`dispatch`; `intake` declared them near the bottom; `herregistratie` declared them
+  after `draftSync`. Three files, three orders for the same three roles.
+- `registratie` declared `draftSync` in the middle of a run of `computed`s.
+- `herregistratie`'s `snapshot` read `this.stepLabels.length` seven lines before
+  `stepLabels` was declared.
+- `intake`'s class comment claimed answers persist to `sessionStorage`. Thirty lines below,
+  another comment said draft-sync replaced it. The class comment was false.
+- All three cited "RD-05's tag-transition rule" in an identical sentence.
+
+## Decisions (pre-made, don't relitigate)
+
+1. **One nine-section member order, identical in all three files.** Injected stores → inputs
+   → the store (`createStore`, `state`, `dispatch`) → static copy → state projections →
+   controllers → shell wiring → event handlers → constructor and `restart`. Each section
+   carries a `// --- <name> ---` header. Section 7's header already existed in all three
+   files, so it is reused byte for byte.
+
+2. **The store ⇄ `draftSync` cycle stays.** The store's effect map calls `this.draftSync`,
+   and `draftSync`'s `snapshot` calls `this.state()`. Both are arrow functions that run after
+   construction, so there is no temporal-dead-zone hazard; the cycle cannot be removed by
+   reordering. One comment at the effect map names it.
+
+3. **Extract `digitalDocumentIds` into `libs/shared/src/domain/upload.machine.ts`.** The
+   "digital and finished uploading" filter was written out four times, in two shapes: mapped
+   to ids in the two container snapshots, and as `.some(...)` inside two machines'
+   `hasProgress`. One function serves all four, and it removes the `r.documentId!` non-null
+   assertion from both containers. It joins an existing file beside `deliveryRefs`, and its
+   spec joins the existing `deliveryRefs` block.
+
+4. **Extract `diplomaMsg` into a new sibling of the step files.** `onDiplomaKeuze` was the
+   fattest member in the three containers and had no test. It is pure: a `DuoLookupDto` and a
+   selected id in, a machine message or `null` out. `HANDMATIG` does not move — the new file
+   is its sibling and imports it exactly as the container did.
+
+5. **`phase` stays in all three containers.** It maps this machine's tags onto the shell's
+   `WizardPhase` vocabulary and composes a `$localize` failure message. That is a container's
+   job. It cannot move to `domain/`: `WizardPhase` comes from an Angular component, and
+   `domain/` points inward only. Moving it to a per-wizard sibling would add three files and
+   three specs to relocate 17 readable lines each. Decision 1 already fixes what was wrong
+   with it — it belongs in the shell-wiring section, and only two of three files had it there.
+
+6. **The four `err*` computeds in `herregistratie` stay.** They are one-line projections
+   feeding four template bindings. Folding them into one `errors()` would edit the template,
+   which is behaviour-shaped work this ticket does not do.
+
+7. **Comment policy, four rules.** Delete ticket references and keep the sentence (RD-18
+   decision 1 stripped `WP-`/`RB-` for the same reason; `RD-` is the same debt). Delete
+   history — a comment says what the code does now. Delete a comment that describes code
+   which is not in the file. Keep a comment that states a current why: the `untracked`
+   loop-avoidance notes, the `IntakePolicy.RejectIncompleteScholing` seam pointer, and the
+   `demo-*` preview note all stay.
+
+## Traps
+
+- **`messages.en.xlf`.** Every `$localize` id in these files is translated. The reorder moves
+  the copy arrays; it must not touch an id or its source text.
+- **`enabled: () => this.seed() === initial`** is a reference-identity check against the
+  exported `initial` singleton. Never clone or rebuild it — breaking the identity turns
+  draft-sync on inside Storybook and the tests.
+- **Public members are load-bearing.** `showcase/vragenlijst.section.ts` reads
+  `IntakeWizardComponent.steps`; `showcase/form-machine.section.ts` reads
+  `HerregistratieWizardComponent.state`. No `private`/`protected`/`readonly` modifier changes.
+- **`intake.machine.ts` carries `#region showcase:steps` markers** that feed `gen:snippets`.
+  This ticket does not touch them.
+- **New specs change a generated document.** `scripts/ci-local.sh` regenerates
+  `libs/shared/docs/behaviour-spec.mdx` and diffs it. Four new `it()` titles land there, so
+  the regenerated file belongs in the same commit.
+
+## Acceptance
+
+- `npm run ci` passes.
+- Each container's member list and every modifier are unchanged:
+  `diff <(git show HEAD:$f | grep -oE '^  (private|protected|readonly)? ?[a-zA-Z]+ *[=(]' | sort) <(...)`
+  reports no difference for all three files.
+- `git grep -nE "\bRD-[0-9]+" -- 'apps/ssp/src/app/*/ui/*wizard*'` returns nothing.
+- `git grep -n "sessionStorage" -- 'apps/ssp/src/app/herregistratie/ui'` returns nothing.