[BE] Stop an assignment showing as tagged when the last competency rule measuring it is archived

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

In short

What changes, and who decides archive vs. delete. delete_competency_criterion() (#674) and delete_competency_criteria_group() (#675) each branch on whether learner mastery exists: the hard-delete branch already reconciles the tag association through tag_object(), removing it when no other criterion still needs it, but the archive branch does nothing to the association at all, which is why an assignment keeps showing as tagged after the last rule measuring it is archived. This ticket makes the archive branch run that same reconciliation call. Which outcome it produces, archive or delete, is not decided here: #782 locks an association against deletion when a learner's mastery is recorded through it, and #779 makes tag_object() retire a locked association instead of deleting it. So calling the same reconciliation from both branches gets the right result for free, and this ticket adds no archive-versus-delete logic of its own.

This ticket adds one check: is the tag still needed at all? Once this operation has finished archiving and hard-deleting whichever criteria it's removing, does any other active criterion still point at this association? If so, leave the association alone. If not, make the same tag_object() call described above to drop the tag. Checking after all the writes, rather than one criterion at a time, is what keeps a mixed subtree correct with no extra code.

Nothing is needed for the re-tagging direction. When an author tags the same assignment again, #779 already makes tag_object() revive the existing retired association in place, automatically. The duplicate-association check in associate_competency_criterion() needs no change either: #717 already scopes it to non-archived criteria, which is what lets the new criterion succeed even though the revived association is the same row the old archived criterion still points at. The work here is a regression test, not code.

Implementation specifics

• Where the change lands. src/openedx_learning/applets/cbe/api.py, in delete_competency_criterion() (#674) and delete_competency_criteria_group() (#675), inside the transaction.atomic() block each function already opens.
• One shared helper, called by both branches. Add a private helper in the same module, for example _reconcile_object_tags(candidate_objecttag_ids: set[int]) -> None, and move #674's existing hard-delete-branch reconciliation onto it so both branches run identical code. Two parallel copies of this rule drift, and the mixed delete-plus-archive case inside a single #675 subtree operation is only reliably correct when one pass evaluates the whole operation after all criteria writes have landed.
• Collect candidates before the criteria writes. Build the candidate set from the oel_tagging_objecttag_id values of the CompetencyCriterion rows the operation is about to archive or delete, read before those rows are modified. A hard-deleted criterion no longer carries that reference afterwards.
• The still-in-use query. CompetencyCriterion.objects.filter(oel_tagging_objecttag_id__in=candidate_objecttag_ids, archived=False).values_list("oel_tagging_objecttag_id", flat=True). Every id it returns is dropped from the candidate set and left entirely untouched. The query is global: it is not filtered by group, subtree, course, or the operation's own scope.
• No ancestor-group walk is needed. Filter on CompetencyCriterion.archived (#716) only. #675 archives every descendant criterion, not just the group row, in the same transaction, so a criterion under an archived group is itself archived and the criterion-level filter is sufficient. CompetencyCriteriaGroup carries no association reference of its own, so groups contribute no candidates.
• Grouping and call shape. Group the surviving candidates by (object_id, taxonomy_id). For each group: values = [t.value for t in get_object_tags(object_id, taxonomy_id=taxonomy_id)], drop that group's affected values, then openedx_tagging.api.tag_object(object_id, taxonomy, values). Read object_id, the Taxonomy object, and the tag value off each ObjectTag before calling tag_object(), because the call may delete the row.
• Scope every call to the association's own taxonomy. Pass taxonomy_id to get_object_tags() and the matching Taxonomy to tag_object(), so no other taxonomy's tags on the same assignment are re-validated or replaced.
• Already-retired associations missing from the reduced list are harmless. #778 excludes retired associations from get_object_tags(), so an assignment that already has a retired association in the same taxonomy will have that value absent from the list handed to tag_object(). Under #779 a locked association absent from the list is retired again, which is a no-op on a row already in that state, and reviving only happens for values that are present in the list, so an omission cannot bring anything back.
• No new code on the create path. associate_competency_criterion() (#665) gets no changes. It must keep creating associations through openedx_tagging.api.tag_object() rather than creating an ObjectTag directly, because only tag_object() revives an existing retired association in place; a direct create would leave a second association alongside the retired one and the assignment would show the tag twice.
• The duplicate check stays exactly as #717 wrote it, scoped to non-archived criteria. Cover it with a regression test rather than changing it.
• Out of scope, assumed already merged. The archived and deletion_locked fields on the tagging models (#776); the lock written when mastery is recorded (#782); tag_object()'s retire-instead-of-delete and revive-on-re-apply behavior (#779); the exclusion of retired associations from ordinary tagging reads and the read exemption that keeps learner mastery traceable (#778); the archived fields on the competency criteria models (#716). Do not reimplement any of these here.
• Layering. openedx_learning sits above openedx_tagging in the src_layering contract in .importlinter, so calling openedx_tagging.api from the CBE applet is the sanctioned direction. Nothing in this ticket may add knowledge of competencies to openedx_tagging.
• Public API impact: none. delete_competency_criterion() and delete_competency_criteria_group() keep their signatures, and the new helper is private. The only behavior change is on the archive branch.
• Tests, in tests/openedx_learning/applets/cbe/test_api.py: archiving the last criterion when learner mastery exists retires the association rather than deleting it, and get_object_tags() stops returning it; archiving one of two criteria that share an association leaves the association untouched and the remaining criterion still evaluates; a group archive spanning two assignments produces exactly one tag_object() call per assignment; a group archive where an active criterion outside the subtree still references one of the assignments leaves that association untouched; a mixed subtree where one criterion is hard-deleted and another is archived, both pointing at the same association, reconciles that association exactly once and retires rather than deletes it; deleting a criterion no learner was measured against still removes the association outright, as a regression guard on #674's existing behavior now routed through the shared helper; re-tagging after an archive revives the same association row rather than creating a second one, creates the new criterion successfully, and leaves the old criterion archived; a learner mastery row for an archived criterion still resolves back to its association.
• ADR alignment, no new ADR needed. This implements the post-use archive rule and the traceability exemption already decided in docs/openedx_learning/decisions/0003-competency-criteria-versioning.rst, Decision 3.

Files to create and modify Modified files

• Context #674 adds delete_competency_criterion(), whose hard-delete branch already reconciles the association through tag_object() and whose archive branch is the gap this ticket closes.
• #675 adds delete_competency_criteria_group(), which archives or deletes a whole subtree of criteria across potentially many assignments in one transaction.
• #665 adds associate_competency_criterion(), the create path this ticket verifies but does not change.
• #776 adds the archived and deletion_locked fields to the tagging models, inert on merge.
• #782 sets the deletion lock on an association whenever learner mastery is recorded against the criterion evaluated through it.
• #779 makes tag_object() retire a locked association instead of deleting it, and revive a retired association in place when the same tag is applied again.
• #778 excludes retired associations from ordinary tagging reads and keeps the learner-mastery traceability read exempt.
• #716 adds archived to the competency criteria models; it is a separate field from the tagging models' own archived and the two should not be conflated.
• #717 scopes the duplicate-association check to non-archived criteria.
• src/openedx_tagging/api.py holds tag_object() and get_object_tags(), the two functions this ticket calls.
• docs/openedx_learning/decisions/0003-competency-criteria-versioning.rst, Decision 3, is the decision record for the post-use archive rule and the traceability exemption.
• .importlinter defines the layering that puts openedx_learning above openedx_tagging.