atomic-design-poc/docs/project/readable-codebase/PLAN.md

Make the rest of the codebase as readable as the dashboard

This is the living design record for the arc, committed so that the repository holds the complete state and a fresh session needs nothing outside it. Corrections found while executing are folded in and marked. See README.md for the ticket ledger and the session protocol.

Context

The dashboard refactor (branch refactor/readable-dashboard, 3 commits, pushed) cut dashboard.page.ts from 340 lines to 42 by splitting it into six per-concern section components. It was built as a reference implementation: prove the pattern on one screen, then hold the rest of the app to the budget it establishes.

This plan applies that result to the rest of the codebase. It is grounded in fresh measurement, not in the original plan's estimates — several of which turned out to be stale (see "Corrections" at the end).

Two findings reframe the work:

1. Pages are already thin. 17 pages, median 88 lines; only the sanctioned showcase teaching page exceeds 250. The "page → sections" split is essentially done. The remaining bulk sits one layer down, in organisms.
2. The worst problem is not size, it is a silent-failure idiom. runIfSubmitting is copy-pasted into 5 components under 2 names and must be called by hand right after dispatch. Forgetting it fails silently. That is a correctness risk, not a cosmetic one.

Sequencing matters, and the phase numbers are not quite the running order:

1. Phase 3a first — the max-lines rule plus reportUnusedDisableDirectives: 'error'. It is one small commit, it lands green today, and it stops every later phase from adding a new oversized file. Cheap insurance bought before the spending starts.
2. Then Phase 0 (put the dashboard in its right context) — small, self-contained, and it settles where the section files live before anything else edits them.
3. Then Phase 1 (idioms) — it deletes code that the Phase 3 splits would otherwise have to carry: the copy-pasted submit plumbing, the WizardStatus switch, several computeds. Split first and you split code you are about to delete.
4. Then Phase 3b–3i (the splits), Phase 2 (mechanical sweeps, independent — fit anywhere), Phase 4 (layer move), Phase 5 (docs) last, so it documents the end state once rather than tracking each step.

---

Execution: how this survives a restart and runs on Sonnet agents

This document is a design record. It is not executable as-is: a fresh Sonnet agent has none of the conversation that produced it, and nothing here records progress. So the first ticket converts it into the artifact this repo already uses for exactly this.

Reuse the existing protocol, do not invent one

docs/project/backlog/README.md is a proven mechanism — 75 work packages driven to done through it, and its own notes say the tickets are "self-contained (each WP file carries its own current-state handoff) and sized for a fresh Sonnet session." Copy it wholesale:

• One ticket file per commit, using that README's existing template (Status / Why / Read first / Decisions (pre-made, don't relitigate) / Files / Steps / Acceptance criteria / Verification / Out of scope / Risks).
• Status: todo | in-progress | done inside each ticket file.
• A README with an Order table carrying every ticket, its dependencies and its status.
• A runnable GREEN one-liner as the global definition of done.

New home: docs/project/readable-codebase/, prefix RD-NN. It must not extend docs/project/backlog/, because Phase 5 archives that directory — a finished arc gets archived, a new arc gets its own folder. RD- also avoids collision with the existing WP-/RB- prefixes, which matters because Phase 2 greps for those.

The one property that makes a restart safe

Each ticket updates its own Status: line and the README row in the same commit as its code. Never in a follow-up commit. That makes git log and the ledger impossible to desync: whatever is committed is done, whatever is not is not.

Recovery for a fresh session with zero context is three commands:

git log --oneline -8
grep -rn '^Status:' docs/project/readable-codebase/RD-*.md | grep -v done   # next work
npm run ci                                                                   # is HEAD green?

Making each ticket Sonnet-executable

An agent reads its own ticket, not this whole document. So each ticket must be self-contained. Ticket files are written just in time by the supervisor, immediately before delegating, and land in that ticket's own commit — not all 35 up front, which would be speculative. Three rules when writing one:

1. Copy the decision, never a pointer to it. The Decisions (pre-made, don't relitigate) block carries the verdict from this plan verbatim. No agent re-derives "effect map vs full Elm" — that is settled here, with reasons, and re-opening it wastes an Opus-shaped judgment on a Sonnet-shaped task.
2. Inline the traps that apply to that ticket. The Risks section below is global; an agent will not read it. The Seed exemption belongs in RD-05's Decisions block, the longest-key-first sed order in RD-27's, the parameterised-$localize rule in RD-25's and RD-26's. A trap left only in a global list is a trap that fires.
3. State acceptance as a command, not a sentence. npm run lint exits non-zero, or the file is under 250 rule-lines, or npx eslint --report-unused-disable-directives is clean. "Lands ~230 lines" is a design estimate and is not checkable; do not put it in Acceptance.

GREEN for this arc

npm run ci

Plus, for any ticket touching a story, an .mdx, or libs/shared/src/ui/**:

npm run ci --full     # the only thing that builds Storybook and catches a broken MDX import

Phase 4's move commit must run --full. So must anything in Phase 3 that moves a template.

The agent loop

One supervisor session drives it; one developer agent (Sonnet) executes each ticket. Per iteration:

1. Read the README Order table. Pick the first todo whose dependencies are all done.
2. Spawn one developer agent with a fixed prompt: "Read CLAUDE.md, then docs/project/readable-codebase/README.md, then RD-NN.md and its Read-first list. Execute it. End GREEN. Update the ticket's Status and the README row in the same commit. Do not start another ticket."
3. Verify with a task-runner agent (Haiku): npm run ci, git log -1 --stat, and that the Status: line now reads done.
4. Green → next iteration. Red → stop and surface. Never mark a ticket done on an agent's word alone; the check is the exit code.

For unattended running, /loop with that iteration as its prompt works — the ledger is the state, so a loop that dies mid-arc resumes from the ledger with nothing lost.

Sequential by default — and why parallel is worse here

Every ticket must end npm run ci green on the branch, and three properties make concurrent writes to one branch actively hostile:

• libs/shared/docs/behaviour-spec.mdx and showcase/snippets.generated.ts are regenerated and drift-checked by CI. Two agents both regenerating conflict by construction.
• Every ticket writes the same README Order row table — contention on literally every iteration.
• The file sets overlap heavily: the three wizards appear in RD-07, RD-08, RD-20, RD-22 and RD-23.

Where parallelism does pay: genuinely disjoint tickets, in separate git worktrees (isolation: "worktree" on the Agent tool), merged deliberately. Good candidates: the ticket sweep (RD-18/RD-19 — 186 files, semantically touching nothing), and the Phase 5 doc tickets. Cap it at two at a time.

Evidence for caution: .claude/worktrees/ currently holds 22 abandoned agent checkouts at 4.7 GB (RD-09 deletes them). Parallel worktree agents have been used in this repo before and left the debris behind. Use them deliberately, and clean up.

The ticket table (RD-01 materialises this verbatim as the README Order table)

Each row becomes one RD-NN-<slug>.md and one commit. "Deps" must all be done before a ticket is picked. The phase sections below this table are the source for each ticket's Decisions block.

ID	Ticket	Deps	Source	--full?
RD-01	Scaffold docs/project/readable-codebase/ — README + all RD files	—	Execution
RD-02	max-lines rule + reportUnusedDisableDirectives: 'error' + 7 disables	01	3a
RD-03	overzicht context: page + 2 nav sections, dep-cruiser edge, HEADER_ADMIN_LINKS	02	0	yes
RD-04	Story titles → Domein/<Context>/<Name>; add stories only where >1 state	03	0	yes
RD-05	createStore gains the effect map + 5 specs	02	1a
RD-06	Bug fix: 2 single-step forms → effect map + retry affordance	05	1a	yes
RD-07	Add Primary to the 3 wizard machines + specs	05	1a
RD-08	Migrate the 3 wizards to the effect map + Primary	07	1a	yes
RD-09	Docs + generator: plop-templates/form-machine.hbs, ARCHITECTURE, fp-tea, skill	08	1c
RD-10	WizardStatus → WizardPhase (payload-carrying)	08	1b#4	yes
RD-11	Fold the projection into remote-data.ts; PascalCase the 3 machines	01	1b#3
RD-12	ActionState → action on BriefState.Loaded	11	1b#2a
RD-13	Same for org-template, folding pendingPublish in	12	1b#2a,#6
RD-14	Move SaveState to debounced-save.ts; delete action-state.ts	13	1b#2b
RD-15	Delete .claude/worktrees/ (22 checkouts, 4.7 GB)	01	2.1
RD-16	parseDashboardView returns BigProfile DROPPED — would discard decisions	01	2.2
RD-17	successOf/successOr sweep — 10 sites, 8 files	01	2.3
RD-18	Ticket sweep, frontend — 181 refs / 100 files	01	2.4
RD-19	Ticket sweep, backend — 370 refs / 86 files	01	2.4
RD-20	wizard-errors.ts + spec, adopted by all 3 wizards	02	3c
RD-21	rich-text-dom.ts helpers + spec cases	02	3h	yes
RD-22	intake-wizard → 3 step components	08, 20	3c	yes
RD-23	registratie-wizard → 3 steps + upload-controller move	08, 20	3c	yes
RD-24	concepts.page → 6 sections + concept-card + 2 globals + --app-code-*	02	3g	yes
RD-25	org-template-editor → sample-letter.ts + labels + 2 children	02	3e, 3f	yes
RD-26	letter-canvas → labels + letter-line; keep its disable	02	3d, 3e	yes
RD-27	The layer move: 33 git mv + 28 specifiers + 8 MDX imports	21	4a	yes
RD-28	Layer-tag fixes (async missing, breadcrumb Chrome) + beheer doc rule	27	4a, 4b
RD-29	The 3 ladder rules in .dependency-cruiser.base.js	27	4c
RD-30	Archive backlog/ + refactor-backlog-setup/ (16,300 lines) + archive README	01	5.1
RD-31	ARCHITECTURE.md §6a — symbols not lines, 2 dead paths, new section names	03, 08, 16	5.2
RD-32	fp-tea-atomic-design.md — 11 broken paths + the broken anchor	27	5.3
RD-33	CLAUDE.md + atomic-design.mdx + ui-component skill	03, 27, 29	5.4-5	yes
RD-34	(optional) NO_SUBORGS/NO_TABLES → RemoteData.Empty	11	1b
RD-35	(optional, last, alone) upload type: → tag:	27	1b#5

Recommended running order is the ID order; it already respects every dependency. RD-15 through RD-19 are independent of everything and can be pulled forward whenever a short session needs filling — RD-15 in particular makes every later search faster and should go early.

Two ordering traps the table encodes but an agent should be told outright:

• RD-01 must precede RD-30, because RD-01 copies its ticket template out of the very directory RD-30 archives.
• Four tickets edit the same two doc files in different sections — RD-09 rewrites the submit-idiom teaching (ARCHITECTURE.md §2d area, fp-tea §338-350), while RD-31 rewrites ARCHITECTURE.md §6a and RD-32 fixes fp-tea's paths. Sequential is fine; never run these two pairs in parallel worktrees.

---

Phase 0 — Put the dashboard in the right context

The dashboard is the portal home, but it lives inside registratie, a context that .dependency-cruiser.ssp.js declares as registratie: [] — permitted to import no other context. Three concrete symptoms:

• Its six sections span four concerns: registratie data (3), aanvragen (1), cross-context action links to /herregistratie /intake /brief /concepts (1), admin links to /beheer/* (1).
• The cross-context coupling is invisible to the linter, because wat-wilt-u-doen.section.ts links by route string. dep:check passes and gives false assurance on exactly this file.
• beheer-links.section.ts:6 imports ADMIN_LINKS from ../../../shell/nav.config — a context reaching into the app frame. app.config.ts:72 already provides that same constant to the shared site header through the HEADER_ADMIN_LINKS token.

Steps

1. Scaffold a context with npm run gen:context (plop context) named overzicht (Dutch, per CLAUDE.md: domain contexts are Dutch; the page's own title is "Mijn overzicht").
2. Move the page and the two portal-level navigation sections into it:
   • overzicht/ui/overzicht.page.ts (was registratie/ui/dashboard.page.ts)
   • overzicht/ui/wat-wilt-u-doen.section.ts
   • overzicht/ui/beheer-links.section.ts
3. Leave the four data sections in registratie/ui/dashboard/ — mijn-aanvragen, wat-moet-ik-regelen, mijn-registratie, specialismen render registratie data and belong beside their store. This separation is only possible because of the split; it is the refactor's first real payoff.
4. Declare the edge in .dependency-cruiser.ssp.js: overzicht: ['registratie'] — the second sanctioned cross-feature edge, mirroring herregistratie: ['registratie'].
5. beheer-links.section.ts injects HEADER_ADMIN_LINKS instead of importing shell/nav.config, removing the context→shell reach-in.
6. Consider giving wat-wilt-u-doen's action list the same treatment — a token beside NAV_ITEMS/ADMIN_LINKS in shell/nav.config. Route strings on a landing page are legitimate, but the list is app-frame copy, not registratie's.
7. Update app.routes.ts:19 to loadComponent: () => import('@overzicht/ui/overzicht.page'), and add the @overzicht/* alias to apps/ssp/tsconfig.json.

Keep the /dashboard route — it is user-visible and in the e2e specs. Renaming it to /overzicht is a separate, optional change needing a redirect.

Also settle the two deviations the refactor left behind:

• Story titles — fix. They are Domein/Registratie/Dashboard/<Name>; CLAUDE.md specifies Domein/<Context>/<Name>, "full stop", and all 41 other story files comply. Retitle to Domein/Registratie/<Name> for the four that stay and Domein/Overzicht/<Name> for those that move. Safe: only layers.mdx deep-links a story id, and not one of these three.
• 8 imports — accept, do not fix. CLAUDE.md has no import-count rule; ≤6 was a proxy metric, and one import per rendered section is exactly right. The only way down is a DASHBOARD_SECTIONS const spread into imports:, which trades a self-documenting array for an indirection and adds a barrel-shaped thing to a repo that deliberately has none.
• Three of six sections have no story (beheer-links, wat-moet-ik-regelen, wat-wilt-u-doen). Add one only where the section has more than one visual state.

---

Phase 1 — One submit idiom, and two state encodings instead of six

This phase fixes two live bugs. It is not only hygiene.

1a. Submit

runIfSubmitting is a private async method duplicated across 5 components — named runIfIndienen in the registratie wizard — invoked at 8 call sites, always as:

this.dispatch({ tag: 'Submit' }); // the reducer decides
this.runIfSubmitting(); // then re-read state() and re-check the tag it hoped for

• behandeling/ui/besluit-form/besluit-form.component.ts:129
• registratie/ui/change-request-form/change-request-form.component.ts:175
• herregistratie/ui/herregistratie-wizard/herregistratie-wizard.component.ts:275
• herregistratie/ui/intake-wizard/intake-wizard.component.ts:386
• registratie/ui/registratie-wizard/registratie-wizard.component.ts:634

The two bugs (fix these regardless of the rest)

besluit-form.component.ts and change-request-form.component.ts never dispatch Retry, although besluit.machine.ts:64 and change-request.machine.ts:50 both export it. On a failed submit, Failed falls into the @else branch that renders the form, but editing() is null there, so besluit()/toelichting() return '' — the user sees an emptied form. The still-enabled submit button dispatches Submit, which the reducer no-ops from Failed. Unrecoverable dead end, reachable from any failed submit, in both components.

Other defects to close in the same pass:

• Step-boundary logic lives in the component (s.step < 3 ? Next : Submit), beside the reducer's own exported next/submit.
• The two herregistratie wizards interleave optimistic store calls the reducer knows nothing about (herregistratie-wizard.component.ts:277,282,285; intake-wizard.component.ts:388,396,399). registratie-wizard has no optimistic flag.

Design: createStore gains an effect map

Constraint: reduce stays pure (CLAUDE.md). The fix lands in the dispatch wrapper in libs/shared/src/application/store.ts — inside the one sanctioned wiring idiom, not beside it.

export type StoreEffects<Model, Msg> = Model extends { tag: string }
  ? {
      [K in Model['tag']]?: (
        state: Extract<Model, { tag: K }>,
        store: Store<Model, Msg>,
      ) => unknown;
    }
  : never;

export function createStore<Model, Msg>(
  init: Model,
  update: (model: Model, msg: Msg) => Model,
  effects?: StoreEffects<Model, Msg>,
): Store<Model, Msg>;

Four load-bearing choices:

• A conditional type, not a Model extends {tag} constraint — the constraint would break store.spec.ts:8's createStore(0, (n, m) => n + m). With the conditional, Model = number resolves to never, so effects are a compile error there and omitting them stays legal.
• Keys are Model['tag'] — a renamed or typo'd state tag becomes a compile error.
• The narrowed state is argument one — this is what deletes the const s = this.state(); if (s.tag !== 'Submitting') return; preamble at all 8 sites. The body cannot run in the wrong state, so it cannot re-guess.
• The store is argument two — the effect needs dispatch, but createStore(...) runs in a field initializer before this.store exists.

The trigger rule is where the design lives. Run effects[next.tag] when both:

1. prev.tag !== next.tag — the store entered the tag. A Submit that fails validation is Editing → Editing: no fire. A second Submit while Submitting is a reducer no-op: no fire, so double-submit protection falls out of the rule. Retry is Failed → Submitting: fires, so onRetry needs no special case.
2. the msg tag is not Seed. This single line is what stops the five Submitting Storybook stories from firing real HTTP (see Risks), and stops draftSync.onResume → Seed from re-submitting a resumed draft.

Implementation note for whoever writes it: capture prev/next inside the model.update(...) callback and invoke the effect after update returns. Do not read model() inside dispatch — the comment at store.ts:24-28 explains the livelock, and store.spec.ts:16-30 exists because that bug already happened once.

Effect bodies do not move. They stay as private component methods, registered in the map; the begin*/confirm*/rollback* calls stay inside them. The effect slot is the sanctioned place for effects, so "side effects stay out of the reducer" holds unchanged. Per-site diff: delete two guard lines, take the narrowed state, register one map entry.

private store = createStore<BesluitState, BesluitMsg>(initial, reduce, {
  Submitting: (s, store) => this.submitBesluit(s, store),
});

onPrimary → a Primary msg on the 3 wizard machines: primary(s) = isLastStep(s) ? submit(s) : next(s), composed from each machine's own already- exported next/submit/currentStep. Keep Next and Submit in the unions (templates, specs and the showcase use them). Then onPrimary() is one dispatch, onRetry() is one dispatch, and the shell's outputs map 1:1 onto messages — which is what .claude/skills/form-machine/SKILL.md:74-78 already claims.

Retry affordance: reuse the existing id @@wizard.opnieuwProberen with byte-identical source text. It already has an English target in both messages.en.xlf files, so no new translation is needed.

Rejected alternatives (recorded, not re-litigated)

• An Angular effect() watching state — reject. store.spec.ts:16-30 exists because this exact pattern livelocked the app. Worse, a signal effect is a latest-value notification, not an event stream: two dispatches in one tick coalesce, so a transient Submitting can be observed as never having happened — a silently dropped submit, the very failure being fixed.
• Full Elm reduce -> [state, Cmd] — the honest end state and the only option with statically exhaustive effect coverage, but disproportionate here: 9 machines, 9 specs, every dispatch site. And domain/ may not import Angular (lint-enforced), so a Cmd cannot carry BigProfileStore.beginHerregistratie — it becomes a symbolic tag plus a switch in the component, which is the guard being deleted, relocated. Keep as the documented upgrade path.

1b. Six encodings of "in flight / ok / failed" → two

Two survive: RemoteData<E,T> for "data I fetched", and the machine's own state union for "where this thing is". Everything else either folds into one of those or is an honest exception with a reason.

#	Encoding	Where	Disposition
1	RemoteData<E,T>	application/remote-data.ts (91 lines), 19 files	Keep, and absorb #3 so one file is where a lifecycle becomes async state
2a	ActionState	application/action-state.ts, 2 producers	Delete — both stores collapse it to busy + lastError byte-identically; zero consumers keep the union. Becomes an action field on each machine's Loaded variant
2b	SaveState	same file	Keep — it has 2 genuine 4-way consumers (brief.page.ts:149-160, org-template.page.ts:102-113). Move it to debounced-save.ts, beside its only producer, then delete action-state.ts
3	LoadLifecycle + machineRemoteData	application/machine-remote-data.ts (24 lines), 3 identical call sites	Relocate, not delete — fold into remote-data.ts as fromLoadLifecycle, keyed PascalCase, beside the existing fromResource
4	WizardStatus	layout/wizard-shell/wizard-shell.component.ts:19	Replace with a payload-carrying WizardPhase — the 3 switches drop the error payload, which then travels as a second errorMessage input
5	UploadStatus	domain/upload.machine.ts:11-18	Keep — it is encoding #2 for a sub-machine, and its payloads (progressPct, documentId, reason) are consumed by 3 UI components. Only the type: dialect is off; optional, last, alone
6	5 ad-hoc flags	see below	1 folds, 4 keep

Why #3 relocates rather than deletes: the mapping has to exist somewhere, because <app-async> takes RemoteData. Deleting the module re-inlines a 6-line switch in 3 stores — recreating the duplication WP-31 removed. Relocating still wins the whole prize: the lowercase constraint dies, the dialect drift resolves, and the survivor sits beside fromResource as what it actually is — a RemoteData constructor, not a sixth encoding.

The causal chain that makes the dialect fix free: brief, org-template and stamdata-editor are the only three state unions in the repo with lowercase tags, their Msg tags are PascalCase in the same file, and the only thing pinning them is machineRemoteData's structural S extends LoadLifecycle. Relocate the projection with PascalCase keys and the drift resolves itself — no separate renaming pass.

#4 — the shell genuinely needs the payload. Replace WizardStatus + errorMessage with one input: { tag:'Editing' } | { tag:'Submitting' } | { tag:'Submitted' } | { tag:'Failed'; message: string }. The 3 mapping computeds stay — Answering and registratie's Dutch Invullen/Indienen/Ingediend/Mislukt are not the shell's vocabulary — but now carry the message, so the 3 duplicated errorMessage computeds fold in and the second input disappears. @switch cannot narrow, so read Failed via the existing whenTag (kernel/fp.ts:27). wizard-shell.stories.ts must change in the same commit.

#6 — one folds, four keep. Only org-template.store.ts:59 pendingPublish is a genuine illegal-state pair (pendingPublish && busy is representable and meaningless) — it becomes a fourth action variant. The other four are correctly modelled as they are: big-profile.store.ts:61 pending is a lone boolean the dashboard reads after the wizard is destroyed, so it cannot be derived from the machine; and aanvragen.store.ts:28, admin-cases.store.ts:26, feature-flags.page.ts:96 are one-shot action errors sitting beside a successfully-loaded list. Folding those into the list's RemoteData would make the error replace the list, since Failure carries no value — precisely the RB-20 behaviour those comments exist to prevent. Add a comment saying so, and leave them.

A win this reveals: org-template.store.ts:29,129 (NO_SUBORGS) and stamdata.store.ts (NO_TABLES) dispatch LoadFailed for what is semantically Empty. Once the projection is explicit, a distinct state → RemoteData.Empty is a few lines, and <app-async> already renders it via emptyText. Optional, own commit, needs one new $localize id.

1c. Commit order

Ticket mapping: A1=RD-05, A2=RD-06, A3=RD-07, A4=RD-08, A5=RD-09, B4=RD-10, B3=RD-11, B2a=RD-12, B2b=RD-13, B2c=RD-14, B6=RD-34, B5=RD-35.

Submit first — it settles the final shape of the 5 components, and 1b#4 touches 3 of them.

#	Commit	Notes
A1	createStore gains the effect map + specs	No call sites change. Behaviour-neutral, so risk is isolated to one commit
A2	Migrate the 2 single-step forms; add the retry affordance	Closes both bugs. Reuses @@wizard.opnieuwProberen, no new xlf target
A3	Add Primary to the 3 wizard machines + specs	Pure domain. check:seam greps BESLUIT_TAGS and SCHOLING_THRESHOLD_DEFAULT — keep both greppable
A4	Migrate the 3 wizards to the effect map + Primary	runIfSubmitting/runIfIndienen become registered effects
A5	Docs and the generator, same diff	See below — non-optional
B4	WizardStatus → WizardPhase	Drops the errorMessage input; update wizard-shell.stories.ts
B3	Fold the projection into remote-data.ts; PascalCase the 3 machines	Widest mechanical commit; 77 lowercase literals
B2a/b	ActionState → action on Loaded (brief, then org-template + pendingPublish)	Must follow B3 or the same tags get renamed twice
B2c	Move SaveState to debounced-save.ts; delete action-state.ts	Type-only
B6	(optional) NO_SUBORGS/NO_TABLES → Empty	Needs a new $localize id per app that renders it
B5	(optional, last, alone) upload type: → tag:	~20 files; see the sed hazard in Risks

A5 — corrected while executing RD-09. The original claim here was that plop-templates/form-machine.hbs generates runIfSubmitting, and that .claude/skills/form-machine/SKILL.md:65-78 teaches it, so that the next scaffolded form would recreate the bug. Both are false, verified by grep:

• plop-templates/form-machine.hbs is machine-only — 74 lines, no @Component, no runIfSubmitting. No generator recreates the bug.
• .claude/skills/form-machine/SKILL.md never mentions it.

Two real sites remain, and both teach the deleted idiom verbatim in a code block: docs/reference/architecture/ARCHITECTURE.md:314 and :574, and docs/reference/fp-tea-atomic-design.md:342. So A5/RD-09 is a two-document fix, still worth doing — a teaching document that teaches a deleted idiom is precisely the rot this arc targets — but it is not the urgent generator fix this plan originally claimed.

Scope discipline: A1 + A2 alone deliver "dispatching cannot silently skip the effect" and fix both bugs. If the budget shrinks, stop after A5; B3 and B2 are hygiene, not correctness.

---

Phase 2 — Mechanical sweeps (no behaviour change)

1. Remove .claude/worktrees/ — 22 abandoned agent-<hex> checkouts, 4.7 GB, gitignored (.gitignore:64). They are why unqualified repo-wide grep/find return ~23× inflated counts, which taxes every future search by a human or an agent.

   Correction made while executing RD-01: these are live registered git worktrees, not orphaned directories. Each has a worktree-agent-<hex> branch carrying real RB-xx commits. So rm -rf is wrong — it leaves 22 broken worktree registrations behind. Use git worktree remove per worktree, then delete each branch, then git worktree prune.

   Verified during RD-01: the commits are already reachable from main (spot-checked 95bb773, 80de261, dfc6c41, ce95294 with git merge-base --is-ancestor), because 637d500 merged the whole RB-01..RB-33 arc. RD-15 must re-verify all 22 before removing any — check every branch tip is an ancestor of main, and stop if one is not.

2. Finish Step 2's name collapse. DROPPED while executing RD-16 — the instruction was wrong, and following it would have introduced a bug.

   This plan claimed DashboardViewDto → DashboardView → BigProfile was "three names for one payload" and that parseDashboardView should return BigProfile directly. Reading the type disproves it:

   export interface DashboardView {
     profile: BigProfile;
     decisions: HerregistratieDecisions;
   }
   

   DashboardView is a pair, and BigProfile is { registration, person } — one member of that pair, with nowhere to put decisions. Returning BigProfile directly would silently discard the server-computed herregistratie eligibility, which is exactly what ADR-0001 says the front end must render rather than recompute.

   The store's two map calls are not a redundant hop either: they project one aggregate into two independently-consumed signals, and six files consume them separately — for example mijn-registratie.section.ts takes profile while wat-moet-ik-regelen.section.ts takes decisions.

   So the three names are a wire DTO, a screen-shaped aggregate, and a component of that aggregate. Three different things, correctly named.

   The other half of Step 2 was already done correctly: HerregistratieDecisions lives in registratie/domain/registration.ts:40, not in contracts/, and only one hand-written contracts file remains (duo-diplomas.dto.ts, a different endpoint). Commit 42e7a1e did the parts that were right and correctly left alone the part that would have been wrong.

3. successOf / successOr sweep — 10 inline unwraps remain in 8 files. They do not all want the same helper:

   • undefined fallback → existing successOf: beoordeling.page.ts:78
   • [] / null fallback → add successOr(rd, fallback): werkvoorraad.page.ts:61, admin-cases.page.ts:83, audit.page.ts:99, feature-flags.store.ts:27, registratie-wizard.component.ts:509, mijn-aanvragen.section.ts:93
   • boolean predicates → leave as-is: access.store.ts:36, feature-flags.store.ts:51
   • big-profile.store.ts:57 hand-rolls map — use the existing map from remote-data.ts

4. Ticket-comment sweep, frontend and backend (user-selected scope): 551 refs across 186 files — 181 in apps+libs (100 files), 370 in backend/ (86 files: 70 .cs, plus Dockerfile, 4 .sh, 4 .yml, 2 .md). Strip WP-/RB- and keep the surrounding sentence; git blame holds the provenance. Keep all 90 ADR-000x refs across 68 files — those point at documents that exist. Note CD- appears nowhere in the repo.

---

Phase 3 — The guard first, then 7 splits

3a. The guard goes FIRST, not last

Putting the rule ahead of the splits stops the refactor itself from adding a new 300-line file. In eslint.config.mjs (which today has no per-folder rules at all):

{
  files: ['{apps,libs}/**/*.{page,component,section,step}.ts'],
  rules: { 'max-lines': ['error', { max: 250, skipBlankLines: true, skipComments: true }] },
},

Two corrections to the approved plan's Step 8, both load-bearing:

• The glob must include section and step. *.{page,component}.ts does not match *.section.ts — the file kind the dashboard refactor invented, and the kind Phase 3 creates most of, would escape the guard entirely.
• Add linterOptions: { reportUnusedDisableDirectives: 'error' } in the same commit. ESLint 9 only warns by default and npm run lint does not fail on warnings. At error, every split commit is forced to delete its own eslint-disable or lint fails — the exemption list cannot rot into permanent debt. The repo has zero disables outside the generated api-client.ts and is clean under this flag today, so it lands green.

3b. Seven offenders, not nine

Measured with the real rule (skipBlankLines, skipComments), not wc -l:

wc -l	rule	File	Axis
644	574	registratie-wizard.component.ts	per step
496	472	showcase/concepts.page.ts	per teaching section
463	414	letter-canvas.component.ts	labels + one extraction, then stays exempt
406	368	intake-wizard.component.ts	per step
353	329	org-template-editor.component.ts	per output cluster
293	253	rich-text-editor.component.ts	over by 3 — move 2 helpers
288	252	herregistratie-wizard.component.ts	over by 2 — one shared helper
267	232	behandel-scherm.component.ts	already compliant — leave alone
262	236	stamdata-table-editor.component.ts	already compliant — leave alone

3c. The step contract already exists — do not invent one

For the wizards, the pattern to copy is not the dashboard. It is registratie/ui/address-fields/address-fields.component.ts, which this very wizard already composes and whose header comment is the contract, verbatim: "Pure & presentational — values in via value, errors in via errors, every keystroke out via fieldChange. No store, no services, no internal state; the container owns the Model and decides what a change means." Two containers already reuse it.

So: inputs down, one narrow output up, dispatch never passed down.

• registratie-wizard → adres.step.ts (~90), beroep.step.ts (~130), controle.step.ts (~110); parent lands ~230. RegistratieLookupStore is providedIn: 'root', so the beroep step injects the same instance and owns its own <app-async> over the DUO lookup — this is the one place the dashboard's axis does apply. Moving the upload controller is what gets the parent under 250: createUploadController takes a dispatch callback, so the step creates its own with dispatch: (msg) => this.uploadMsg.emit(msg) — one output, not five. The BRP prefill effect stays in the parent, because it writes to the machine.
• intake-wizard → buitenland.step.ts, werk.step.ts, review.step.ts; parent ~200. scholingZichtbaar is not an input — each step takes the threshold and calls the pure lageUren(answers, threshold) itself ("derive, don't store").
• herregistratie-wizard is over by two lines. Do not split its steps for symmetry — its whole template is ~100 lines. Extract one shared pure helper instead: layout/wizard-shell/wizard-errors.ts with toWizardErrors() + spec, beside the existing naarStapLabel that lives there for exactly this reason. Removes ~6 lines from all three wizards. Do not touch the three shellStatus switches — the tags genuinely differ per machine, and an exhaustive switch is the house style.
• wizard-shell (205 lines) already provides the whole frame — stepper, error summary, <form>, navigation, submitting/submitted/failed, a11y focus. The steps slot into its existing default <ng-content>. Nothing new in libs/shared.

Corollary: give the new steps no stories. Each wizard's existing story already mounts every step by seeding the machine. The dashboard got this right too — 3 stories for 6 sections, only where there was async state to show.

3d. letter-canvas — a misapplied rule, not a split

20 of its 28 input()s are pure $localize labels and no caller overrides a single one across all 4 call sites. The CLAUDE.md rule they were built for — "Shared/English components must not hardcode Dutch — expose copy as input()s" — governs libs/shared, not a Dutch domain component in brief/ui/. Inline them as i18n="@@id" in the template; same id, same source text means zero messages.en.xlf edits.

Do not collapse them into a config object or an injection token. HEADER_NAV_ITEMS/DEBUG_PANEL exist because two apps genuinely differ; here nothing differs, so a token adds a provider and an indirection to solve a problem nobody has.

One extraction earns its keep: letter-line.component.ts (~70 lines out) — the #line template plus the sample/diff helpers. It replaces six 4-line ngTemplateOutlet incantations with three one-line tags and is the only part with logic worth a spec.

That leaves ~329: 77 lines of CSS and 204 lines of one letter. Splitting it into letterhead/body/signature/footer makes "what does the letter look like" a five-file question for no behavioural seam. Keep one /* eslint-disable max-lines */ with an honest reason. It becomes the only disable in the repo, and reportUnusedDisableDirectives keeps it honest.

3e. The $localize boundary that governs 3d and 3f

Plain messages move to the template; parameterised ones stay in TS. The xlf stores interpolations as <x id="min" equiv-text="MARGIN_MIN_MM"/>; moving such a message into a template renames the placeholder to INTERPOLATION and breaks the translation merge, so ng build --localize fails. Only 4 messages are affected: orgTemplate.margins, orgTemplate.invalid, orgTemplate.publish.impact, wizard.naarStap.

3f. org-template-editor — split by output cluster

The 11 output()s are the tell; each child takes one mutation family:

• SAMPLE_LETTER_BRIEF (44 lines) → brief/domain/sample-letter.ts. It is a dead export (used only in its own file) and it is production content, not a fixture — so it must not go near brief.testing.ts, or dependency-cruiser's no-testing-in-production rule fails.
• 11 of 13 label inputs → template i18n (they are declared protected, so they were never bindable — constants wearing input() ceremony). The two parameterised ones stay per 3e.
• logo-upload.component.ts (~34 out) and version-history.component.ts (~18 out).
• Parent drops 11 outputs → 5 and lands ~222. No exemption.

3g. concepts.page.ts — split by section, but decompose the CSS by owner

A per-section split does not fix the 142-line styles: block, because Angular scopes styles per component: the page's .card cannot style a child's DOM. So:

• .section → delete, use the existing global .app-section.
• .lead, .cols → two new globals beside .app-text-subtle/.app-stack in libs/shared/styles.scss, whose own comment says it exists to centralise these idioms.
• .card, .tag*, .note, pre (~70 lines) → owned once by concept-card.component.ts, used ~11 times, which also deletes ~11 copies of the card boilerplate.

Watch the colour guard. scripts/check-tokens.sh greps only --include='*.component.ts' — which is why this page currently gets away with #1e2430, #fff, #e5e5e5. Moving that CSS into a *.component.ts brings it under the guard for the first time, so in the same commit: drop the var(--rhc-x, #hex) fallbacks (all 16 tokens are defined) and add --app-code-bg/-fg/-keyword/-string/-comment for the pre palette, following the existing --app-devpanel-* precedent added for this exact reason.

3h. rich-text-editor — over by three

rich-text-dom.ts already exists beside it with its own spec, so the seam is built. Move the selection/range surgery out of deleteAdjacentChip/insert into it. Cheapest of the seven, and it converts two untested imperative-DOM branches into spec cases.

3i. Order

1. the rule + reportUnusedDisableDirectives + 7 disables
2. wizard-errors.ts + spec, adopted by all three wizards → delete that disable
3. rich-text-dom helpers + spec → delete that disable
4. intake-wizard → 3 steps
5. registratie-wizard → 3 steps + the upload-controller move
6. concepts.page → 6 sections + concept-card + 2 globals + --app-code-* tokens
7. org-template-editor → sample-letter.ts + labels + 2 children
8. letter-canvas → labels + letter-line; keep its disable, rewrite the reason

Steps 2–8 are independent; only 2 must precede 4 and 5.

---

Phase 4 — Make the folder equal the layer, then enforce the ladder

4a. Move (libs/shared/src/ui/ only)

ui/atoms/       12 flat + upload/{delivery-channel-toggle,document-chip,file-input,
                              upload-progress-bar,upload-status-icon}        (17)
ui/molecules/   13 flat + upload/single-upload                              (14)
ui/organisms/   upload/{document-category,document-upload}                    (2)

upload/ splits by layer but keeps its feature subfolder inside each layer. It satisfies decision #2 literally, costs the same 6 relative-import rewrites as a flat split, keeps a genuinely cohesive group together, and makes "5 atoms + 1 molecule + 2 organisms" visible in the tree instead of hidden in story titles. No barrel — the repo has none and does not need one.

layout/ does not move. CLAUDE.md §5 explicitly enumerates libs/shared/layout components getting Atoms…Templates buckets, so layout/ is sanctioned to hold several layers; its organisms are chrome only its own templates compose. Instead, fix the two mislabels: async.component.ts is missing its /** Molecule: */ tag, and breadcrumb.component.ts says /** Chrome: */ where its title says Molecules.

libs/beheer/src/ui/ — fix the doc, not the code. Its story title is Domein/Beheer/Stamdata Table Editor while CLAUDE.md §5 and layers.mdx say Design System/…. The code is right: libs/beheer is a bounded context that lives in libs/ only to be shared by two apps. Amend the two doc lines. That dissolves the "two taxonomies" oddity and beheer correctly needs no layer folders.

Mechanics, one commit for all of ui/: 33 whole-directory git mvs break zero relative imports except the 6 inside upload/; then rewrite the 28 distinct specifier strings (179 occurrences, 59 files) longest-key-first, so upload/<sub>/ is processed before any bare upload/. Confirmed: no edits to angular.json, eslint.config.mjs, plopfile.mjs, e2e/, or any tsconfig; both Storybook globs are recursive; dependency-cruiser's ui-not-infrastructure pattern still matches a nested path; and check-tokens.sh's CIBG-GAP check keys on the directory basename, which a parent-folder move preserves.

Verify with npm run typecheck (4 tsconfigs — catches every missed specifier), dep:check, test, then npm run ci --full. git diff --stat -M should show only renames plus one-line import edits.

4b. Keep all 78 layer-tag comments

Reversal of my earlier claim. The tag prefixes a real one-line description (/** Atom: thin wrapper over CIBG .btn — typed variant API. */); deleting the word leaves the sentence and buys nothing. And only the 32 in ui/ are made redundant by folders — the 25 organisms and 9 pages in apps/**/ui/ have no layer folder and a title that deliberately omits the layer, so there the comment is the sole carrier. A three-way redundancy that has never once disagreed is cheap documentation.

4c. The ladder rules are the real prize

The folder move is what makes this expressible; this is what makes it enforced. Today nothing stops an atom importing an organism. Add to .dependency-cruiser.base.js:

• atoms-compose-nothing-above: ui/atoms/ → ui/(molecules|organisms)/ forbidden
• molecules-below-organisms: ui/molecules/ → ui/organisms/ forbidden
• design-system-not-layout: ui/ → layout/ forbidden

Two details that matter: forbid upward only, never "atoms are leaves" — same-layer edges are legitimate and exist today (masked-value → button, review-section → data-block, task-list → choice-link); and pathNot must exempt \.(spec|stories)\.ts$, because async.stories.ts composes skeleton and a story may legitimately reach for context.

Zero upward edges exist today, so all three land green immediately. Dependency-cruiser rather than ESLint: it is where every other boundary rule lives and it emits the architecture graph.

---

Phase 5 — Fix the docs that describe this flow

1. Archive the finished backlog. git mv docs/project/backlog and docs/project/refactor-backlog-setup under docs/project/archive/. Verified: all 74 WP files are Status: done; the two trees are 6,982 + 9,318 = 16,300 of the docs tree's 20,317 lines. Add a ~15-line archive/README.md: this is historical, git holds the rest.
2. ARCHITECTURE.md §6a ("The request lifecycle today", line 542) is the best onboarding artifact in the repo and has rotted:
   • 15 L<n> line citations, now wrong — Program.cs L80 lands on a // WP-60: comment about client timeouts, not /dashboard-view; L120 lands mid-expression. Cite symbols, not lines.
   • Two cited paths do not exist: src/environments/environment.ts (pre-monorepo) and proxy.conf.json.
   • Its read-walkthrough shows <app-async [data]="store.profile()"> on the dashboard page; after the split that markup lives in mijn-registratie.section.ts.
   • It states the boundary yields RemoteData<DashboardView> — wrong once Phase 2.2 lands.
3. docs/reference/fp-tea-atomic-design.md — 11 pre-monorepo src/app/… paths, all broken; one (submit-herregistratie.ts) points at a deleted file; and a broken anchor at line 427 (#1-the-big-picture-three-contexts-four-layers vs the actual "two apps, cross-app libraries").
4. Update CLAUDE.md for the new overzicht context, the max-lines budget, the libs/beheer title rule (4a), and the step-component contract (3c).
5. libs/shared/docs/atomic-design.mdx gains the step contract and the layer table; also fix its stale claim that eslint.config.mjs enforces the layer rules — dependency-cruiser does. Fix the stale src/app/shared/ui/... paths in .claude/skills/ui-component/SKILL.md.

---

Risks

1. The Storybook trap (highest). All 5 components mount their Submitting/Indienen state via Seed in stories that use a real provideHttpClient() with no request mocking (besluit-form.stories.ts:33, herregistratie-wizard.stories.ts:70, intake-wizard.stories.ts:38, change-request-form.stories.ts:34, registratie-wizard.stories.ts:88). Without the Seed exemption in the trigger rule they fire real network calls, flip to Failed, and red the storybook-a11y job (npm run ci --full). Verifying those 5 stories still show a spinner is the exemption's acceptance test.
2. dispatch becomes effectful, and two dispatch sites sit inside Angular effect()s — intake-wizard.component.ts:363 (SetPolicy) and registratie-wizard.component.ts:~596 (PrefillAdres). Both land on an unchanged tag, so nothing fires, and both are already untracked. Never key an effect on the editing tag — that is the livelock.
3. The type: → tag: sed hazard (B5). type is also a legitimate field name in that neighbourhood — rejectReason(cat, { type: file.type, sizeMb }) — and FileRejected.reason has the literal value 'type'. A blind rename breaks upload validation silently. File-by-file with the type-checker, or defer.
4. The ticket sweep is 186 files, including backend/'s Dockerfile, 4 .sh and 4 .yml. Do not blanket-sed. A WP-/RB- token could appear in a string that matters (seed data, a test name, a migration id) rather than a comment. Review the diff per file group, and keep all 90 ADR-000x refs.
5. CI drift gates bite mechanically. scripts/ci-local.sh:33-34 diffs showcase/snippets.generated.ts (fed by // #region showcase: markers — remote-data.ts:30 carries showcase:fold, intake.machine.ts:59 carries showcase:steps) and libs/shared/docs/behaviour-spec.mdx. Run gen:snippets / gen:behaviour-spec in the same commit as any change that moves a region or a spec title.
6. action inside Loaded narrows where "busy" can exist. Re-check every current actionState.set({tag:'Busy'}) site for reachability from a non-loaded state. Note a BriefLoaded reload then resets action to Idle, which is a behaviour improvement: a stale error can no longer outlive a reload.
7. Highest risk in Phase 4: a broken MDX story import passes npm run ci and fails CI. libs/shared/docs/*.mdx has 8 relative story imports across a11y.mdx, atomic-design.mdx, remote-data.mdx and fp-in-ui.mdx of the form ../src/ui/<name>/<name>.stories. These break on the move, and only build-storybook catches them — which is not in the default npm run ci, only --full. Fix all 8 in the move commit and run npm run ci --full before pushing it.
8. check-tokens.sh has blind spots in both directions. It greps only --include='*.component.ts'. Moving CSS from a *.page.ts into a *.component.ts newly exposes it (see 3g); moving CSS to a .styles.ts or .scss newly hides it. Decide deliberately, and widen the glob in the same commit if you move CSS out.
9. styles: [importedConst] is unproven in this repo — 47 of 47 components use inline literals and only 2 .scss files exist. If the letter-canvas style extraction is attempted, prove it with ng build first; the fallback is the single eslint-disable.
10. Behaviour drift while moving 200+ template lines. Move template text byte-identically and let git diff -M prove it. The wizards' specs cover reduce, not the markup, so the markup's only guards are review and the axe run in ci --full.
11. A step component reaching for the store. Passing dispatch down is tempting and would let a step dispatch Submit. Use outputs, per address-fields. If overridden, type the input as Extract<RegistratieMsg, {tag: 'SetField' | …}> so an illegal dispatch is unrepresentable.

Verification

Per commit:

1. npm run ci — lint (incl. the new max-lines), typecheck, dep:check, format, tokens, both apps' tests, ng build --localize (catches a missing messages.en.xlf target for any moved $localize string), audit, backend dotnet test, API-client drift.
2. cd backend && dotnet test for anything touching the wire.

End to end, after Phase 0 and Phase 3:

3. npm start, open http://localhost:4200/dashboard. All six sections render as before.
4. ?scenario=loading, =empty, =error, =slow — each section shows its own state, not one page-wide spinner.
5. ?role=admin — the Beheer section appears; without it, absent.
6. Resume and cancel a concept aanvraag — optimistic update and error path both work.
7. Walk each wizard end to end after Phase 1: submit, retry after a failure, and the step-boundary transitions.
8. npm run storybook and npm run storybook:behandelportal — moved and new stories render, a11y addon clean.

---

Corrections to the original plan's claims

Measured against the current tree, not assumed:

• Step 2 did not fully land. This correction was itself wrong, and is withdrawn. The chain is intact because it should be: DashboardView is a pair of BigProfile and HerregistratieDecisions, not a third name for either. Collapsing it would discard the server-computed decisions. The claim was made by reading the parse signature without reading the type it returns. See Phase 2.2, now dropped.
• 7 files exceed 250 lines, not 8. The plan counted by wc -l; the rule as specified uses skipBlankLines + skipComments. concepts.page.ts (472) was missing from its list, but behandel-scherm (232) and stamdata-table-editor (236) were on it and already pass.
• 551 ticket refs across 186 files, not 478/170. CD- does not exist anywhere.
• Only 2 of the 4 named adapters make no HTTP call. letter-preview.adapter.ts and reveal-bignummer.adapter.ts both fetch for real (deliberately hand-written, not the generated client). Only MedewerkerAdapter and DigidAdapter are pure stand-ins, and both already carry a // ponytail: fake … label. Renaming those two is cosmetic — low priority.
• registratie-wizard's Dutch tags (Invullen/Indienen/Ingediend/Mislukt) are correct, not drift — CLAUDE.md requires Dutch domain contexts. Do not "fix" them.
• Storybook titles and header comments agree in all but two cases across the 41 story files, so the folder=layer move is mechanical, not a taxonomy debate. The two: async.component.ts has no tag at all, and breadcrumb.component.ts says /** Chrome: */.

A live CI-gate defect, found while executing RD-06 (now fixed). scripts/ci-local.sh gated its storybook + axe steps on [[ "${1:-}" == "--full" ]], but CLAUDE.md documents npm run ci --full — and npm parses that flag itself, exporting npm_config_full=true instead of passing --full through as $1. Proven with npm run env --full. So the documented command skipped both steps and still printed "local CI passed": a gate reporting success without running. The script now accepts either form, which makes every existing doc correct rather than requiring them all to change. This matters directly for RD-27, whose highest risk is a broken .mdx story import that only build-storybook catches.

And four corrections to claims made earlier in this same investigation, caught by reading the consumers and the rule semantics rather than the definitions:

• "~62 of the 77 layer-tag comments become deletable" — withdrawn. Keep all 78 (the count was also one short). The tag prefixes a real description, so deleting the word leaves the sentence; and only the 32 in libs/shared/src/ui/ are made redundant by folders. Elsewhere the comment is the sole carrier of the layer.

• libs/shared/src/ui/ holds 17 atoms, not 16.

• SaveState is not redundant. It has 2 genuine consumers that keep all four cases (brief.page.ts:149-160, org-template.page.ts:102-113). Only ActionState collapses, so action-state.ts must be split, not deleted.

• "Five ad-hoc boolean/nullable-string pairs" was overstated. Only org-template.store.ts:59 is a genuine illegal-state pair. Three of the five (aanvragen.store.ts:28, admin-cases.store.ts:26, feature-flags.page.ts:96) are a lone signal<string | null> with no boolean partner — an action error beside a loaded list, on a different axis from the fetch, correctly modelled as it stands.

Deliberately out of scope

• Renaming the /dashboard route to /overzicht (needs a redirect; user-visible).
• Splitting libs/shared/src/layout/ by layer (see 4a for why not). Cost if you disagree: 10 distinct specifiers, 34 occurrences, 28 files.
• Splitting herregistratie-wizard's three steps for symmetry with the other two wizards. It is over budget by two lines and its template is ~100 lines. Worth noting as a consistency follow-up, not as work.
• behandel-scherm.component.ts and stamdata-table-editor.component.ts — both already pass the rule. Leave them alone.
• The 7 non-component files over 250 lines (brief.adapter.ts 408, upload.machine.spec.ts 364, brief.store.spec.ts 357, …). The glob deliberately does not reach them.
• A scripts/check-layers.sh asserting folder == tag == story title. Cheap (~15 lines, modelled on the CIBG-GAP check) but it only catches doc typos.
• The admin Case/Zaak vocabulary rename — a separate read model.
• NgRx, real auth, runtime DTO validation on every endpoint (CLAUDE.md "out of scope").