[FE] Show and remove competency criteria whose gradeable subsection has been deleted

This section is background and a suggested approach, not the ticket's source of truth. The User
Story and Acceptance Criteria define what must be true when the work is done.

In short

Nothing on the backend hides these criteria, so no backend change is needed. The tagging row
a criterion points at stores the subsection as a plain opaque string with no foreign key to course
content and no resolver, and criterion creation validates only that the string parses as a usage
key whose course matches the group's scope. Deleting the subsection leaves every one of those rows
untouched, so the endpoint that returns a competency's groups and criteria returns an orphaned
criterion like any other. The one backend requirement is a negative one, that this read path never
acquires a filter on whether the content still exists, which is recorded in Open Questions rather
than built here.

How the page decides a criterion is unresolvable. Subsection titles are resolved on the client
by taking the criterion's usage key, reading the outline of the course that key names, and finding
the matching node. A criterion counts as unresolvable only when that outline was read successfully
and contains no node with that usage key. Every other state is not evidence of a deletion: a
request still in flight, a request that failed, and a course the author has no access to must each
render as their own state. Treating them as deletions would show authors phantom warnings whenever
a request failed.

Which course outlines the page loads, and through which query. Parse the course key out of
each criterion's usage key, reduce that to the distinct set of courses the competency's criteria
reference, and read one outline per course. The page needs exactly these outlines anyway to title
the criteria that do resolve, so this adds no request that titling does not already require. Read
them through the same cached query the course browse panel uses rather than a second request path,
so a course already expanded in the browse panel costs nothing to resolve here.

Removal needs no new endpoint. The single-criterion removal endpoint (#674) takes a criterion
id, decides internally between deleting the row and archiving it depending on whether a learner
status row references it, and never touches the shared tagging row. None of that depends on the
subsection existing, so an unresolvable criterion removes exactly like any other one.

What the criterion shows in place of a title. Name the state in words rather than leaning on an
icon or a color, and include the course the deleted subsection belonged to, so an author with
several unresolvable criteria can tell which course each one came from and judge which competency
rule is affected.

Implementation specifics

• Extend the Selected Content components #672 adds under
  src/taxonomy/competency-management/criteria-groups/, rather than introducing a parallel
  rendering path for these criteria.
• Derive the course key with getCourseKey from src/generic/key-utils.ts. It throws on a
  string it cannot parse, so wrap the call: a criterion whose stored key does not parse as a course
  block key is unresolvable with no course to name, and must not break the surrounding list.
• Read outlines through the query the course browse panel already uses.
  useCourseOutlineIndex (src/course-outline/data/outlineIndexQuery.ts) wraps
  getCourseOutlineIndex and keys on courseOutlineQueryKeys.index(courseId). The number of
  courses is only known at render time, so a hook call per course is not possible: build the batch
  with useQueries over that same key factory and query function, following the batching precedent
  in src/content-tags-drawer/data/apiHooks.ts. Reusing the key factory is what makes these
  entries share the browse panel's cache instead of duplicating its requests. Carry the hook's own
  retry: false and pass refetchOnMount: false, matching how the browse panel reads outlines.
• Search the returned structure's full section-to-subsection tree for the usage key, starting
  at courseStructure on the response envelope (CourseOutline, src/course-outline/data/types.ts)
  and recursing through each node's childInfo.children; the node fields come from XBlockBase and
  XBlock in src/data/types.ts. Do not assume a position or depth, and do not match on title.
• Branch on the query state explicitly. Only a successful outline read with no matching usage
  key yields the unresolvable state; pending and error states each render their own state, and a
  permission failure is an error, not a deletion.
• Do not block the criteria list on the outline queries. Render the groups and their criteria
  first with titles pending, then settle each criterion as its course's outline resolves.
• Scope a failed outline read to the criteria that depend on it, the way the course browse panel
  renders a per-course outline failure as an inline error on that course's row rather than failing
  the whole panel. One unreadable course must not take down the rest of the competency's criteria.
• Leave the removal control enabled while a title lookup is pending. Removal needs only the
  criterion id, which the criteria response already carries, so it does not depend on the outline
  queries and must not be gated on them.
• Memoize the distinct course-key list on the criteria query's own data reference, not on an
  array recomputed inline in the same render. A dependency rebuilt every render defeats the
  memoization silently, with no type error or failing test to catch it, and here it would refire a
  batch of outline queries.
• Use Paragon's Warning icon from @openedx/paragon/icons alongside the text label, not in
  place of it, and do not distinguish the state by color alone.
• Carry the accessible text in a visually hidden sr-only sibling span, not in an aria-label
  on the wrapper. A plain span or div takes the implicit generic role, which prohibits
  aria-label and is not reliably announced. CompetencyTreeItem.tsx in the same directory already
  does this for the tree's own accessible labels.
• Style the state by overriding Paragon's own design tokens, not the resolved CSS properties.
  Trace the token's resolution in the installed theme first: if it already resolves to the intended
  value, the override is dead code and should not be added. Use a relative unit (rem, %) for any
  typography token, as Paragon's own defaults do, even if a design spec names a pixel value.
• Prefer Paragon's grid for the small-screen behavior over new media queries, as the
  associations panel does with Row/Col. If a component-local SCSS file is added and does use
  Paragon's custom-media breakpoint names, add
  @use "@openedx/paragon/styles/css/core/custom-media-breakpoints.css" in that file: a component
  SCSS imported from its own .tsx compiles as its own root stylesheet and does not inherit the
  @use in src/index.scss, and an unresolved custom-media name is not a syntax error, so the rule
  compiles and then silently never matches. The lint and unit-test commands do not compile Sass, so
  verify any new at-rule by test-compiling the file exactly as the bundler imports it.
• Key any selector for a criterion inside a nested group off the class marking the group unit
  (for example .criteria-group > ul), not a fixed structural depth (> ul > li > ul), which
  matches one nesting level and stops applying past it.
• Do not add a backend existence flag to the criteria read endpoint. That endpoint lives in
  openedx-core, which must not import from openedx-platform (enforced by .importlinter), so
  it cannot resolve a usage key against the modulestore or the content search index. Resolution
  stays on the client, which is also where titling already happens.
• Reuse the confirmation the criteria removal flow already shows. Do not add a second dialog or
  write new confirmation copy for this case.
• A removed criterion that carries learner status is archived rather than deleted, and archived
  records are excluded from read paths, so the criterion leaves the list on either branch without
  this ticket special-casing the outcome.
• Tests: a criterion whose usage key is absent from a successfully read outline renders the
  unresolvable label, the warning icon, and its course; a criterion present in the outline renders
  its title unchanged; a pending outline query renders a loading state and not the unresolvable
  label; a failed outline query and a 403 each render the could-not-check state and not the
  unresolvable label; a group whose every criterion is unresolvable renders all of them; an
  unparseable stored key renders unresolvable without a course and without throwing; removal posts
  the criterion id to the removal endpoint and drops the criterion from the list; several criteria
  in one course fire a single outline query rather than one per criterion, and re-rendering fires no
  further ones; the unresolvable state is present as text in the DOM; the label and course stay
  within the layout at a small viewport width.
• Out of scope: the learner-facing progress surface, which faces the same situation and has its
  own design work ahead of it; the warning shown before deleting a subsection on the Course Outline
  page; any job that prunes or reports orphaned criteria in bulk; and any competency-level
  indication that a competency's rules can no longer be satisfied.

src/taxonomy/competency-management/ holds the Competency Management page, its competency tree,
and its shared messages.ts (#680), with the associations panel and course browse panel in its
associations/ and course-search/ subdirectories (#670). The Selected Content components this
ticket extends are added by #672 in the same tree; the rows below extend them rather than replacing
them.

New files

Modified files

• #672 lists a group's criteria in its Selected Content area and owns the components this ticket
  extends.
• #670 provides the associations panel and the course browse panel, and establishes reading a
  course outline lazily per course through useCourseOutlineIndex.
• #680 provides the Competency Management page, its competency tree, and the page's messages.ts.
• #674 removes a single criterion by id, branching between deletion and archiving.
• src/generic/key-utils.ts for getCourseKey; src/course-outline/data/outlineIndexQuery.ts for
  useCourseOutlineIndex and src/course-outline/data/queryKeys.ts for the key factory to batch
  over; src/course-outline/data/types.ts and src/data/types.ts for the outline envelope and its
  node fields; src/content-tags-drawer/data/apiHooks.ts for the useQueries batching precedent.
• src/taxonomy/competency-management/CompetencyTreeItem.tsx for the sr-only accessible-label
  precedent this page already follows, and src/taxonomy/competency-management/messages.ts for
  where its copy lives.
• .claude/frontend-app-authoring-architecture-overview.md for the authoring frontend's conventions
  and how it consumes this library's APIs.