# Seller Management — Backend Migration Plan

Documentation only. No code, no implementation. This plan synthesizes the
three prior audits — do not re-derive facts already established there,
read them for detail:

- [Seller-Management.md](Seller-Management.md) — capability overview,
  Implemented/Planned/Future legend
- [Seller-Management-Backoffice-Readiness-Audit.md](Seller-Management-Backoffice-Readiness-Audit.md)
  — per-admin-module scoping facts
- [Seller-Management-Storefront-Audit.md](Seller-Management-Storefront-Audit.md)
  — storefront/subdomain facts
- `BACKEND.md` §11 — the backend documentation this plan assumes as its
  starting contract

**Non-negotiable constraint repeated from the mission, and load-bearing for
every classification below:** existing marketplaces without sellers must
continue working exactly as today. Seller Management stays optional
forever, not just at launch. No endpoint may require seller support unless
`modules.sellerManagement.enabled` is true. Every classification and every
migration strategy in this document is written to satisfy that constraint
first — where a module can't satisfy it without a structural rework, that's
called out explicitly, not glossed over.

## Endpoint classification legend

- **No change** — endpoint's request/response/behavior is untouched.
- **Minor change** — an optional field/parameter added (e.g. `sellerId?` in
  a filters object or DTO); absent value behaves identically to today.
- **Major change** — data model or endpoint semantics change beyond an
  optional field (e.g. a new join, a new required decision like split
  orders, a new aggregation dimension).
- **New endpoint** — doesn't exist today, net-new surface.

## Module-by-module

### Authentication
- **Current:** Two live mechanisms (`BACKEND.md` §4) — Telegram/QR
  customer session login (live), Ed25519 admin challenge/response (built
  client-side, not wired live). No seller concept in either.
- **Future:** `SellerPermissionRole` (`marketplaceOwner`/`seller`/
  `sellerStaff`/`platformAdmin`) as a role a session can carry, and a
  decision on whether Seller/Seller Staff authenticate via a third
  mechanism or reuse Telegram/Ed25519 with a different claim.
- **Migration strategy:** additive only — a new optional claim/role field
  on the existing JWT structure (§4 §3), never a new auth mechanism forced
  onto existing flows. Existing customer and admin login must not gain any
  new required step.
- **Backward compatibility:** total, if done as an additive claim. A
  session with no seller claim today is indistinguishable from a session
  before this work existed.
- **Risk:** **Medium** — auth is the most consequence-sensitive surface in
  the app (ADR-010 froze it once already); any change here needs
  security review, not just a schema add.
- **Effort:** Medium (claim/role addition) once the role-vs-`AdminRole`
  reconciliation decision (§11.4) is made; that decision itself is the
  bigger unknown, not the wiring.
- **Endpoint classification:** existing login/verify/refresh/logout —
  **Minor change** (optional claim), contingent on the role decision above.
  New seller-specific login/enrollment flow (if Seller/Seller Staff need
  one) — **New endpoint**, Future, not designed.

### Authorization
- **Current:** Coarse `AdminRole` (Owner/Manager/Support/ReadOnly) mapped
  to `backoffice.read`/`backoffice.write`/`builder.read`/`builder.write`/
  `settings.manage` (§4 §9.1). No route currently enforces even this — the
  admin role model exists but nothing gates on it yet (confirmed by the
  Backoffice audit: "no admin module anywhere does role-based hiding of
  buttons or data today").
- **Future:** `SellerPermissionRole` layered in, deciding whether it
  extends or sits alongside `AdminRole`.
- **Migration strategy:** since authorization enforcement doesn't exist yet
  even for the current roles, there is no existing enforcement to break —
  this is genuinely additive design space, not a migration of live
  behavior.
- **Backward compatibility:** trivial today (nothing to preserve because
  nothing enforces roles yet) but becomes a real compatibility question the
  moment `AdminRole` enforcement is eventually added — sequence matters:
  whichever role model ships first should be designed with the other in
  mind, or the second one becomes a breaking migration of the first.
- **Risk:** **Low today, Medium if sequenced wrong** — the risk is entirely
  about doing `AdminRole` enforcement and `SellerPermissionRole` design in
  the wrong order, not about either individually.
- **Effort:** Low to design (no live behavior to preserve); Medium to
  implement once both role systems' relationship is decided.
- **Endpoint classification:** all admin endpoints — **No change** today
  (no enforcement exists to change); **Major change** whenever
  role-enforcement is added generally, seller-aware or not.

### Bootstrap
- **Current:** `GET /bootstrap`, backend resolves tenant by Host (ADR-001,
  §1). `BootstrapConfig` already has optional `modules?`/`seller?` fields
  (§11.6), always absent/false — no backend populates them.
- **Future:** backend resolving `{seller}.{marketplace-domain}` (§11.5) and
  populating `modules.sellerManagement.enabled` + `seller` for a real
  seller-scoped request.
- **Migration strategy:** the fields already exist and are optional — the
  only backend work is *populating* them correctly for a resolved seller
  scope, not adding new frontend-facing shape. Existing marketplaces'
  bootstrap responses need zero changes; they simply never populate the
  new fields.
- **Backward compatibility:** already verified — these are optional fields
  added to a live contract with zero consumer changes required (confirmed
  `tsc --noEmit` clean at the time they were added).
- **Risk:** **Low** — this is the best-prepared module in the whole
  migration, precisely because the frontend contract was already extended
  in advance without needing backend cooperation.
- **Effort:** Medium — the resolution logic (subdomain → seller lookup) is
  new backend work, but the response shape is already spoken for.
- **Endpoint classification:** `GET /bootstrap` — **Minor change** (two new
  optional response fields to populate, conditionally).

### Products
- **Current:** No real backend yet (§8) — mock gateway. `AdminProductListFilters`
  already has `search/categoryId/visibility/stock/includeArchived/sort/page/pageSize`.
  `Item`/`AdminProduct` already carry optional `sellerId?` (§11.8),
  read/written by nobody today.
- **Future:** `sellerId` filter on list endpoints; ownership enforcement on
  create/update/delete once a real seller session exists.
- **Migration strategy:** add `sellerId` as one more optional field to the
  existing filters object and DTOs — the field already exists in the
  frontend type, so there's no frontend change required at all, only
  backend query logic (`WHERE seller_id = ? OR seller_id IS NULL` pattern
  or equivalent) gated behind the feature flag.
- **Backward compatibility:** guaranteed at the frontend type level
  already; backend guarantee depends on making the new column/filter
  nullable and defaulting existing rows to `NULL` (marketplace-owned) —
  see Database changes below.
- **Risk:** **Low** for the filter addition; **Medium** for ownership
  *enforcement* (a bug here could hide a marketplace owner's own products
  from themselves, or leak a seller's products to another seller).
- **Effort:** Small (list/filter) once the DI-token seam this domain
  currently lacks (per Backoffice audit) is added — that seam is
  independent prerequisite work, not seller-specific.
- **Endpoint classification:** list/get — **Minor change**. Create/update
  (ownership assignment) — **Minor change** if `sellerId` is just an
  optional write field; **Major change** if ownership transfer or
  seller-side write restrictions are added.

### Categories
- **Current:** The most backend-mature domain — real HTTP already
  (`AdminCategoriesApiGateway`, DI-token-bound, §3.2). Facade currently
  hardcodes a full unfiltered fetch and does all real filtering
  client-side to preserve tree parent/child chains (Backoffice audit
  finding).
- **Future:** categories are conceptually marketplace-wide (shared
  taxonomy) — the real future question is scoping *products within* a
  category by seller, not scoping categories themselves.
- **Migration strategy:** likely **no backend change at all** for
  Categories proper; ownership scoping happens one level down, at Products.
- **Backward compatibility:** trivial — no change anticipated.
- **Risk:** **Low.**
- **Effort:** Low (frontend-only fix: the facade's hardcoded full-fetch
  pattern, unrelated to backend).
- **Endpoint classification:** **No change** anticipated.

### Orders
- **Current:** `AdminOrderListFilters` has `search/status/page/pageSize`;
  `AdminOrder` has optional `sellerId?` (§11.8) at the order level.
  `AdminOrderItem` (per-line-item) has **no seller attribution field at
  all** — the concrete gap behind Checkout Modes (§11.7).
- **Future:** per-item seller attribution, plus the Unified-vs-Split-Orders
  decision (§11.7) — the single most consequential undecided item in this
  entire plan.
- **Migration strategy:** cannot be resolved by an additive field alone.
  **Unified Order** path: add optional `sellerId` to each order-item row
  (minor, additive). **Split Orders** path: checkout must decide, at
  creation time, whether to write N order records instead of one for a
  multi-seller cart — that changes `POST /orders`' semantics for any cart
  containing mixed-seller items, which is a **major** change regardless of
  how carefully it's gated, because "how many order records does this
  checkout produce" is not optional-field-shaped.
- **Backward compatibility:** guaranteed for any cart containing only
  marketplace-owned items (the only kind that exists today) under either
  path. The compatibility risk is entirely about *new* multi-seller carts,
  which cannot exist until Products have real seller ownership — so there
  is a natural sequencing safety net here, not just a promise.
- **Risk:** **High** — this is the highest-risk item in the whole plan.
  Payments, refunds, and financial reporting all depend on the
  Unified-vs-Split decision; getting it wrong after sellers exist means a
  breaking change to live financial data, not a code refactor.
- **Effort:** Large — this decision should be made and locked before any
  seller onboarding is possible, not discovered mid-rollout.
- **Endpoint classification:** `GET /orders` (list/filter) — **Minor
  change**. `POST /orders` (create) — **Major change** (semantics change
  based on the Unified/Split decision). `PATCH` status transitions — likely
  **Minor change** if orders stay 1:1 with a single seller scope (Split
  path) or **Major change** if partial per-seller status exists within one
  unified order.

### Payments
- **Current:** **Frozen** (ADR-010) — Telegram QR/card payment flow,
  `POST /cart` (`CartPaymentRequest` → `QrCreateResponse`), status polling.
  Explicitly preserved unchanged through every prior platform-refactoring
  pass in this codebase's history.
- **Future:** if Split Orders is chosen (§ Orders above), a single checkout
  may need to create multiple payment records or split a captured payment
  across sellers — a genuinely new payment-flow shape, not a parameter
  addition.
- **Migration strategy:** **do not touch the frozen flow for the Unified
  Order path** — a unified order with mixed-seller items can still use
  today's exact single-payment flow, only the *order* record differs
  internally. Only the Split Orders path forces any payment-flow change,
  and even then, the existing flow should remain the path for
  seller-disabled marketplaces with zero exceptions.
- **Backward compatibility:** absolute requirement, restated from ADR-010 —
  this is the one area of the whole codebase with an explicit prior
  "freeze, do not redesign" decision, and Seller Management does not
  override it.
- **Risk:** **High** if Split Orders is chosen — payment/financial flows
  are the least forgiving place for a design misstep in this entire system.
- **Effort:** Large, only if Split Orders is chosen; **zero** for Unified
  Order.
- **Endpoint classification:** **No change** under Unified Order. **Major
  change or New endpoint** under Split Orders (undecided, Future).

### Transactions
- **Current:** Derived entirely from the same unscoped Orders list
  (`AdminTransactionsLocalGateway` mirrors `AdminOrdersLocalGateway`'s
  pattern, Backoffice audit finding). `AdminTransactionListFilters` has
  `search/status/type/page/pageSize`.
- **Future:** per-seller transaction filtering/reporting.
- **Migration strategy:** inherits whatever Orders decides — Transactions
  cannot be scoped correctly until Orders resolves per-item seller
  attribution (§ Orders above). Adding a `sellerId` filter here today would
  be cosmetic without that prerequisite.
- **Backward compatibility:** same guarantee as Orders — blocked on the
  same prerequisite, not an independent risk.
- **Risk:** **Medium** — inherits Orders' risk one level removed (financial
  reporting, not the payment flow itself).
- **Effort:** Small once Orders is resolved; not independently schedulable
  before that.
- **Endpoint classification:** **Minor change** (filter field), contingent
  on Orders' Major change landing first.

### Reviews
- **Current:** Storefront submission is live; admin moderation
  (`AdminModerationGateway`) is mock-only. `AdminReviewListFilters` has
  `search/status/rating/page/pageSize`. Reviews carry `productId` — no
  direct seller field.
- **Future:** a seller sees reviews on their own products, via a join
  through Products' `sellerId`, not a new field on the review itself.
- **Migration strategy:** add `sellerId` as a filter that internally joins
  through the product's ownership — additive at the API surface, but
  implemented as a join, not a stored column on reviews.
- **Backward compatibility:** total — the filter is purely additive and
  optional.
- **Risk:** **Low.**
- **Effort:** Small, contingent on Products having real `sellerId` data to
  join against.
- **Endpoint classification:** review list/detail — **Minor change**.
  Reports (`loadReports()`, currently bare no-arg, no filters object at
  all) — **Minor change** too, but requires adding a filters parameter
  that doesn't exist on the endpoint today, so slightly more surface than
  the reviews list.

### Analytics
- **Current:** No aggregation endpoint exists at all (§3.20) — every
  number (revenue, top products, health, recommendations) is computed
  client-side by summing the *entire* orders/products/categories/reviews
  corpus. This gap exists independent of Seller Management.
- **Future:** per-seller analytics, requiring the underlying domains
  (Orders, Products, Reviews) to be seller-scoped first, then a real
  aggregation endpoint partitioned by seller.
- **Migration strategy:** do not attempt to seller-scope analytics before
  a real aggregation endpoint exists for the marketplace as a whole — that
  is a prerequisite gap, not a Seller Management task. Once it exists,
  seller-partitioning is an additional dimension on the same aggregation
  query, not a new endpoint family.
- **Backward compatibility:** unaffected — a marketplace with no sellers
  gets marketplace-wide aggregates exactly as it always has, whenever the
  real aggregation endpoint is eventually built.
- **Risk:** **Medium** — mostly the risk of building the wrong aggregation
  shape before seller-partitioning is even a consideration, and having to
  redo it.
- **Effort:** Large — this is explicitly the last domain recommended for
  real backend work in `BACKEND.md` §9, and Seller Management adds a
  further dimension on top of an already-large lift.
- **Endpoint classification:** **New endpoint** (the aggregation endpoint
  itself doesn't exist yet, seller-aware or not).

### Media
- **Current:** `MediaRepository` abstract class, DI-token-bound already
  (mock live, `ApiMediaRepository` not yet written). `MediaListParams`
  already accepts `page/pageSize/search/folder/kind/sort`. `MediaAsset` has
  no owner/uploader field at all.
- **Future:** `sellerId` filter, requiring an owner field added to
  `MediaAsset` first.
- **Migration strategy:** two small additive changes — a new optional
  `sellerId`/`ownerId` field on the asset model, and a matching optional
  filter field on `MediaListParams`. This module and Categories are the
  two best-positioned in the entire audit for a low-risk addition.
- **Backward compatibility:** total — both changes are optional-field
  additions to an already-flexible params object.
- **Risk:** **Low.**
- **Effort:** Small.
- **Endpoint classification:** list — **Minor change**. Upload/update
  (§7) — **Minor change** (one new optional field in the request body).

### Search
- **Current:** Standard search contract exists (§2.5) as a framework
  convention (keyword, no dedicated backend endpoint beyond catalog list
  filtering — `/search` reuses the same catalog container/endpoint as
  `/catalog` per the Storefront audit). No autocomplete/trending endpoint
  exists; explicitly flagged as its own open "Requires backend decision"
  in §2.12.
- **Future:** filtering search results by seller scope, same mechanism as
  Products (search is not architecturally distinct from catalog browsing
  today).
- **Migration strategy:** inherits Products' migration entirely — no
  separate search-specific backend work anticipated beyond what Products
  already needs.
- **Backward compatibility:** total, same guarantee as Products.
- **Risk:** **Low.**
- **Effort:** None beyond Products — not independently schedulable.
- **Endpoint classification:** **Minor change**, identical to Products'
  classification (same underlying endpoint).

### CMS (Static Pages)
- **Current:** No backend call at all — reads/writes
  `BootstrapConfig.staticPages` in-memory (§1, §3.8). Marketplace-wide
  content (legal, about) by nature.
- **Future:** **conceptually does not apply** — no per-seller static page
  concept exists in this document or in product intent. If sellers ever
  need their own content pages, that is new product surface, not an
  extension of this module.
- **Migration strategy:** none anticipated. Flag if product strategy later
  decides otherwise — treat as a new capability, not a CMS migration.
- **Backward compatibility:** unaffected — no change proposed.
- **Risk:** **None.**
- **Effort:** **None.**
- **Endpoint classification:** **No change.**

### Builder
- **Current:** No backend write path at all (§1.10, §8) — draft/publish is
  `localStorage`-only, editing one global `BootstrapConfig` document per
  marketplace. The single strongest one-owner assumption in the codebase
  (Backoffice audit finding).
- **Future:** a seller-scoped builder (per-seller storefront layout/
  branding) would be an entirely new product surface built on a different
  premise than "one config document per marketplace" — not an extension of
  the existing builder.
- **Migration strategy:** none proposed for the existing Builder. If a
  seller-facing builder is ever wanted, it should be scoped as its own ADR
  and its own backend surface, not bolted onto the marketplace Builder's
  existing draft/publish contract.
- **Backward compatibility:** unaffected — no change proposed to the
  existing module.
- **Risk:** **None** for the existing module; **High** if a future team
  attempts to retrofit seller-awareness into the existing single-document
  model rather than building fresh — flagged explicitly as an
  anti-pattern to avoid.
- **Effort:** **None** now; a seller-facing builder, if ever built, is a
  large, independent effort, not a migration of this one.
- **Endpoint classification:** **No change** to the existing (still
  nonexistent) builder write path. Any future seller-builder surface is
  **New endpoint**, Future, not designed.

### Settings
- **Current:** No route exists — `comingSoon: true` in the admin nav, no
  component backs it (confirmed, Backoffice audit).
- **Future:** if Settings is ever built, it's the natural home for
  marketplace-level Seller Management configuration (e.g. enabling the
  module, default seller policies) — speculative, not designed.
- **Migration strategy:** none — there's nothing to migrate.
- **Backward compatibility:** not applicable.
- **Risk:** **None.**
- **Effort:** **None** attributable to Seller Management specifically.
- **Endpoint classification:** **No change** — no endpoint exists.

### Notifications
- **Current:** A marketplace-facing feature flag exists
  (`FeatureFlagsConfig.notifications: boolean`) but this gates a storefront
  UI feature, not a backend notification/email service — no such service
  exists in this codebase for anyone today.
- **Future:** seller onboarding (invitation accepted, application
  approved/rejected) would need real notification delivery — entirely
  net-new, not an extension of the existing flag.
- **Migration strategy:** none — nothing exists to migrate.
- **Backward compatibility:** not applicable.
- **Risk:** **Low** (net-new work carries normal build risk, not migration
  risk).
- **Effort:** Medium, whenever built — but zero today and not a
  prerequisite for anything else in this plan.
- **Endpoint classification:** **New endpoint**, Future, not designed.

### Emails
- **Current:** No email-sending capability exists anywhere in this
  codebase.
- **Future:** the mocked "Request Access" form on the Phase 1 UI
  (`admin-seller-management-page.component.ts`) implies a real email would
  eventually be sent on submission — today it's a `setTimeout`-mocked
  success dialog with zero delivery.
- **Migration strategy:** none — nothing exists to migrate.
- **Backward compatibility:** not applicable.
- **Risk:** **Low** (net-new, not migration).
- **Effort:** Small to Medium, whenever built (transactional email for one
  form submission is a contained scope).
- **Endpoint classification:** **New endpoint**, Future, not designed.

### Audit Logs
- **Current:** Documented in `BACKEND.md` §5.14 as a proposed convention
  (structured audit-log entries with actor/action/target/timestamp) — no
  backend implementation confirmed to exist; this is itself already
  Planned/Future in the base backend doc, independent of Seller
  Management.
- **Future:** any seller-specific action (seller created, activated,
  suspended, branding changed) should flow through the same audit-log
  convention once it exists, with an added seller-scope dimension on the
  log entry — not a parallel logging system.
- **Migration strategy:** none specific to Seller Management beyond
  ensuring, whenever audit logging is built, that its schema includes an
  optional seller-scope field from day one rather than retrofitting it
  later.
- **Backward compatibility:** not applicable — nothing exists yet to
  preserve.
- **Risk:** **Low.**
- **Effort:** None additional if sequenced correctly (add the field when
  audit logging is first built); Medium if audit logging ships first
  without it and needs a schema migration later.
- **Endpoint classification:** **New endpoint** (audit log query surface,
  if one is ever exposed to admins) — Future, not designed.

## Database changes

No schema exists yet for any of this (§8 — no backend implemented at all).
Recommendations, not decisions:

- A `sellers` table (id, marketplace_id, name, slug, status, branding
  JSON/columns, timestamps) — new table, no impact on existing schema.
- A nullable `seller_id` foreign key on `products`/`items` and `orders` (or
  order-items, pending the Unified/Split decision) — nullable by
  design, defaulting existing rows to `NULL` (marketplace-owned). This is
  the one schema change that touches existing tables; every other addition
  is a new table.
- A nullable `seller_id`/`owner_id` on the media-assets table, if Media
  ownership scoping is pursued.
- No column should ever be added as `NOT NULL` without a default — that
  would force a backfill migration on existing data, which this plan's
  constraint explicitly rules out.

## Permission changes

- New `sellers` and `seller_users` (or equivalent) authorization checks —
  entirely new policy, not a modification of existing `AdminRole` checks
  (which, per the Backoffice audit, don't enforce anything yet regardless).
- Existing admin/customer authorization paths: **no change** required for
  marketplaces with `modules.sellerManagement.enabled = false`.
- Recommendation: implement seller-scoped authorization as an additional
  policy layer evaluated only when a request resolves to a seller scope
  (§11.5) — never as a modification of the existing marketplace-level
  policy evaluation path.

## Caching

- Bootstrap responses are already cached client-side per tenant (§1.7); a
  seller-scoped bootstrap response should be cached under a cache key that
  includes the resolved seller identity (e.g. keyed by full resolved Host,
  which it already effectively is), not the marketplace alone — otherwise
  a seller subdomain risks serving a cached marketplace-level response or
  vice versa.
- No existing cache invalidation logic needs to change for marketplaces
  without sellers.

## Indexes

- `sellers(marketplace_id)` — every seller lookup is scoped by marketplace.
- `products(seller_id)` / `orders(seller_id)` (or order-items equivalent)
  — nullable-column indexes; most databases index `NULL` efficiently, but
  this should be verified against the chosen database engine before
  assuming query performance is unaffected for the common (`NULL`) case.
- No index changes required on any table for marketplaces that never
  populate `seller_id`.

## Security

- Seller-scoped data access is a new cross-tenant-adjacent boundary (a
  seller must never see another seller's data within the same
  marketplace) — recommend treating this with the same rigor as tenant
  isolation (§5.10/§4 §10), not as a lesser internal boundary.
- The existing frozen payment flow (ADR-010) must not be reopened for the
  Unified Order path — only the Split Orders path (if chosen) touches
  payment security surface at all, and that touch should get the same
  security review rigor as the original payment implementation.

## Performance

- Analytics is already the heaviest computation path in the app (full
  client-side aggregation over the entire corpus, §3.20) — building the
  real aggregation endpoint this plan's Analytics section calls for should
  happen with seller-partitioning in mind from the start, to avoid a
  second heavy migration shortly after the first.
- No performance regression is anticipated for marketplaces without
  sellers — every proposed change is either a new table/endpoint (zero
  cost when unused) or a nullable-field filter (negligible cost when the
  filter is never applied).

## API Versioning

Already documented as an open, undecided item in `BACKEND.md` §2.10 (no
path/header versioning scheme exists today). Recommendation specific to
this migration: whichever versioning decision is made generally should
land **before** any Seller Management endpoint ships, so new endpoints
(Seller CRUD, Activation, Invitations, Branding, Analytics, Dashboard) are
versioned consistently with the rest of the API from their first day,
rather than being retrofitted later as the one inconsistent set.

## Migration order

Strict dependency order, not a preference:

1. **Bootstrap** (already prepared, lowest risk) — backend starts
   populating `modules`/`seller` fields for a resolved seller scope.
2. **Products** — `seller_id` column + filter, the foundation every
   downstream domain depends on.
3. **Categories** — confirm no change needed (likely true, low effort to
   verify).
4. **Media** — independent, can happen in parallel with Products.
5. **Reviews** — depends on Products (join through product ownership).
6. **Orders** — the Unified-vs-Split-Orders decision must be made here,
   informed by Products already existing. This is the hard gate — nothing
   past this point should start before it's resolved.
7. **Payments** — only touched if Split Orders is chosen; otherwise
   untouched, in parallel with everything above.
8. **Transactions** — depends on Orders.
9. **Analytics** — depends on Orders, Products, Reviews all being scoped;
   also depends on the pre-existing (non-seller-specific) real aggregation
   endpoint being built first.
10. **Authentication/Authorization** — the `SellerPermissionRole` design
    and its relationship to `AdminRole` should be settled in parallel with
    steps 2-6, not deferred to the end, since Seller Activation and any
    real seller-facing endpoint depend on it.
11. **Seller CRUD/Activation/Invitations/Branding** — depends on
    Authentication/Authorization being settled.
12. **Seller Dashboard/Analytics endpoints** — last, depends on Analytics'
    real aggregation endpoint existing.
13. **Notifications/Emails/Audit Logs** — can start any time after step 11
    (Seller Invitations existing gives them something to notify about);
    not blocking anything else.
14. **CMS, Builder, Settings** — no migration anticipated; revisit only if
    product strategy changes.

## Recommended implementation phases

- **Phase A — Foundation** (steps 1-4, 10 above): bootstrap population,
  Products/Categories/Media scoping, and the permission-role design done
  in parallel. Nothing customer-visible yet. Lowest risk, unblocks
  everything else.
- **Phase B — The hard decision** (step 6, Orders): Unified-vs-Split-Orders
  locked before any real seller can transact. This phase is a design
  decision plus its implementation, not a feature — treat it as a gate,
  not a sprint item.
- **Phase C — Dependent domains** (steps 5, 7, 8): Reviews, Payments (if
  Split chosen), Transactions — mechanical once Phase B lands.
- **Phase D — Seller-facing surface** (step 11): Seller CRUD, Activation,
  Invitations, Branding — the first point where a seller account can
  actually exist and do something.
- **Phase E — Analytics & reporting** (steps 9, 12): the largest single
  remaining lift, deliberately last since it depends on everything above.
- **Phase F — Operational polish** (step 13): Notifications, Emails, Audit
  Logs — improves the experience of Phase D/E but blocks nothing.

At every phase boundary: re-verify the non-negotiable constraint. A
marketplace with `modules.sellerManagement.enabled = false` must show zero
behavioral difference before and after each phase ships. If a phase can't
satisfy that, it isn't ready to ship — regardless of how much of it is
"done."