feat: /v1/role list, get, create, update and delete. Reassign members on delete. Write-time validation (ValidateGrants, duplicates, typed keys against real content types). Name uniqueness (role_name_taken), role_locked, role_has_members, role_conflict on version mismatch. No-escalation: nobody grants a permission they don't hold, and roles.edit_owner_only guards the Owner role. Activity events for role changes. Typed-key exposure in /auth/me. Cleanup of typed rows when a content type is deleted. | Backend only. Will carry the deferred ListGrants decision.

docs/architecture.md

@@ -113,9 +113,9 @@ At boot, `BackfillImageMetadataUseCase` (through `image.Repository.ListMissingMe
 
 ### Roles and permissions
 
-`internal/domain/role` holds the permission catalogue, the built-in roles, one exported constant per catalogue key (`roleKeys.go`, the only place a key is spelled in `api/`) and the pure rules over grants: `ValidateGrants`, `ParseTypedKey` and `Role.GrantSet()`. `Role.GrantSet()` is the single constructor of a `GrantSet` and the only place a stored grant becomes an effective one: it drops unknown keys, typed keys on non-content permissions or with an invalid type key, and owner-only permissions on a role that is not locked; it treats a toggle stored at `own` as `none`; and it resolves duplicate rows to the most restrictive scope. `ValidateGrants` rejects the same rows at write time (`ErrDuplicateGrant`, `ErrOwnerOnlyGrant`, `ErrInvalidScope`, `ErrUnknownPermission`), so a bad row is refused when written and ignored when read. A toggle permission only counts at scope `all`, a scoped one at `own` or `all`. A content permission can be narrowed or widened per content type with a key such as `content.update:blog`; the typed row replaces the wildcard for that type and a `none` row denies. A `GrantSet` is opaque: `ScopeFor(key, typeKey)` resolves wildcard and typed rows, `Permits(key, typeKey, callerID, createdBy)` is the one implementation of own and all (`created_by` 0 and caller 0 own nothing), and `Holds`, `HoldsFor` and `HoldsAll` answer the gate questions. Use cases, handlers and middleware never compare a caller id with `CreatedBy` and never read grant rows. `role.Repository.Save` is version checked: it updates the row only when the caller's version matches, bumps the version, replaces the grants in the same transaction and returns `ErrRoleConflict` otherwise. It never changes `built_in` or `locked`.
+`internal/domain/role` holds the permission catalogue, the built-in roles, one exported constant per catalogue key (`roleKeys.go`, the only place a key is spelled in `api/`) and the pure rules over grants: `ValidateGrants`, `ParseTypedKey` and `Role.GrantSet()`. `Role.GrantSet()` is the single constructor of a `GrantSet` and the only place a stored grant becomes an effective one: it drops unknown keys, typed keys on non-content permissions or with an invalid type key, and owner-only permissions on a role that is not locked; it treats a toggle stored at `own` as `none`; and it resolves duplicate rows to the most restrictive scope. `ValidateGrants` rejects the same rows at write time (`ErrDuplicateGrant`, `ErrOwnerOnlyGrant`, `ErrInvalidScope`, `ErrUnknownPermission`), so a bad row is refused when written and ignored when read. A toggle permission only counts at scope `all`, a scoped one at `own` or `all`. A content permission can be narrowed or widened per content type with a key such as `content.update:blog`; the typed row replaces the wildcard for that type and a `none` row denies. A `GrantSet` is opaque: `ScopeFor(key, typeKey)` resolves wildcard and typed rows, `Permits(key, typeKey, callerID, createdBy)` is the one implementation of own and all (`created_by` 0 and caller 0 own nothing), and `Holds`, `HoldsFor` and `HoldsAll` answer the gate questions. Use cases, handlers and middleware never compare a caller id with `CreatedBy` and never read grant rows. `role.Repository` holds `List` (with every role's grants, in one pass), `Get`, `GetByName`, `Create`, `Save`, `Delete`, `ReassignMembers` and `RemoveTypedGrants`; there is no separate grant listing. `Save` is version checked: it updates the row only when the caller's version matches, bumps the version, writes `updated_at` and `updated_by`, replaces the grants in the same transaction and returns `ErrRoleConflict` otherwise. It never changes `built_in` or `locked`. `Delete` removes the role and its grants in one transaction and returns `ErrRoleHasMembers` when a user still has the role id; `ReassignMembers(from, to)` changes only `role_id` and returns the count; `RemoveTypedGrants(typeKey)` removes the five `content.*:{typeKey}` rows from every role and bumps the version of each affected role once. The role domain also owns `GrantSet.Covers(other, typeKeys)`, the single implementation of "the caller holds at least these grants" behind the no-escalation rule (it returns an `EscalationError` naming the first offending key), `Role.EffectiveGrants()` (the normalised list used by role responses, the audit diff and `/v1/auth/me`) and `DiffGrants`. The roles API (`/v1/role`) is served by `ListRolesUseCase`, `GetRoleUseCase`, `CreateRoleUseCase`, `UpdateRoleUseCase` (which also resets a built-in role), `DeleteRoleUseCase` and `ReassignRoleMembersUseCase`; they depend on `role.Repository`, `user.Repository`, `contenttype.Repository` and `audit.Recorder` only, and `api/handlers/role` maps their domain errors to the status and code. `DeleteContentTypeUseCase` and `CreateContentTypeUseCase` also take the role repository to clear per-type grants. `DeleteContentTypeUseCase` deletes the type and then clears its typed rows; if that cleanup fails the rows stay, which is safe because the role views (`GET /v1/role`, `GET /v1/role/{id}` and every write response) drop typed rows for content types that no longer exist, so a GET body stays a valid PUT body, and `CreateContentTypeUseCase` clears the rows of a key right after a successful create, before any role can match the new type. A failed create never touches the roles. The repository re-checks the name case-insensitively inside the `Create` and `Save` transactions (`ErrRoleNameTaken`), locks the role row in `Delete` before counting members (a no-op on SQLite) and guards the `ReassignMembers` update with an `EXISTS` on the target role, answering `ErrInvalidReassign` when the target vanished.
 
-Enforcement path. The first middleware that needs the signed-in user (`AuditContext`, `RequireCurrentToken`, `RequireVerifiedUser` or a permission gate) stores a request-scoped holder in the context (`access.WithCurrentUser`), so the user is loaded once per request; an API key adds the one load of its own authentication. `CheckPermissionUseCase` resolves the caller's role with one `role.Repository.Get` (or `GetByName` for the legacy owner and admin strings), builds the `GrantSet` from the grants the role already carries and never calls `ListGrants`. `RequireTypedPermission` gates a content route on the typed scope before any lookup (the type key is `article` for articles and the `{key}` of the URL for entries) and puts the caller (`access.Caller`: user id and grants) in the context. The optional `OptionalAccess` middleware does the same for the public article routes without ever refusing: no cookie, an unverified user, a stale token or an unresolvable role leave the caller anonymous, and only a role repository failure answers 500. The handler passes the caller to the use case as `useCases.ContentCaller`; the use case loads the article or entry, decides the scope with `Permits` right after the load and before any other lookup or validation, and returns a `role.MissingPermissionError`, which `access.WriteIfMissingPermission` writes as 403 with the base key in `X-Missing-Permission`. Reading follows the same helper: an article is visible when it is online or `content.read` permits its `created_by`, and an entry the caller may not read is `entry.ErrEntryNotFound`. A list under scope `own` is filtered before totals, search and paging. `CreateContentTypeUseCase` refuses the built-in type keys with `contenttype.ErrKeyTaken`.
+Enforcement path. The first middleware that needs the signed-in user (`AuditContext`, `RequireCurrentToken`, `RequireVerifiedUser` or a permission gate) stores a request-scoped holder in the context (`access.WithCurrentUser`), so the user is loaded once per request; an API key adds the one load of its own authentication. `CheckPermissionUseCase` resolves the caller's role with one `role.Repository.Get` (or `GetByName` for the legacy owner and admin strings), builds the `GrantSet` from the grants the role already carries and never calls `List`. `RequireTypedPermission` gates a content route on the typed scope before any lookup (the type key is `article` for articles and the `{key}` of the URL for entries) and puts the caller (`access.Caller`: user id and grants) in the context. The optional `OptionalAccess` middleware does the same for the public article routes without ever refusing: no cookie, an unverified user, a stale token or an unresolvable role leave the caller anonymous, and only a role repository failure answers 500. The handler passes the caller to the use case as `useCases.ContentCaller`; the use case loads the article or entry, decides the scope with `Permits` right after the load and before any other lookup or validation, and returns a `role.MissingPermissionError`, which `access.WriteIfMissingPermission` writes as 403 with the base key in `X-Missing-Permission`. Reading follows the same helper: an article is visible when it is online or `content.read` permits its `created_by`, and an entry the caller may not read is `entry.ErrEntryNotFound`. A list under scope `own` is filtered before totals, search and paging. `CreateContentTypeUseCase` refuses the built-in type keys with `contenttype.ErrKeyTaken`.
 
 Built-in reconciliation: `database.SeedBuiltInRoles` creates the three built-in roles when they are missing and never overwrites an edited one. On every start it adds the catalogue keys a built-in role lacks: always for Owner, at scope `all`, and for Admin and Writer only while their version is still 1. It never removes or downgrades a grant and never bumps the version.