feat(concepts): full-page browser/detail, links, neighbourhood, robust ontology save

[larsgeorge-db]

Summary

Rebuilds the Concepts experience around a full-page master/detail flow and hardens the backend so LLM-generated ontologies always import cleanly.

Concepts UX

• New Concept detail page at /concepts/browser/:iri with a sticky header (back, title/type/IRI, edit/delete), single dense column matching other detail views. Legacy ?concept=IRI URLs auto-redirect.
• List view is now full-width with richer rows (type icon, badge, status pill, collection chip, property hints) and persisted UI state (expanded groups, scroll position, search) via glossary-preferences-store.
• LinkedObjectsPanel: collapsible card matching NodeLinksPanel style. Assign dialog now restricted to Data Product, Data Contract, and Asset — no raw UC objects. Asset list pulls from /api/assets (paginated) and is bounded + scrollable. Existing UC links keep rendering for back-compat but can no longer be created from this UI.
• ConceptRelationsPanel replaces NodeLinksPanel on detail; consumes /api/semantic-models/neighbors so the full set of incoming/outgoing triples renders, with directional icons and colour-coded predicate badges.
• ConceptNeighborhoodGraph: ReactFlow mini-graph (parents → current → children) with click-to-navigate and theme-correct colours for light/dark.
• Cache-coherence store (knowledge-graph-store): bumped on Settings → Rebuild Graph and on every semantic-link mutation so dependent panels refetch without a hard reload.
• i18n: linkedObjects / neighborhood / relations keys plus the new Asset label across de/en/es/fr/it/ja/nl.

Backend hardening

• clean_truncated_turtle helper in owl/owl_parser.py salvages truncated LLM Turtle; applied in SemanticModelsManager, the base ontology handler, and the upload route. rdflib.parse() failures now surface as HTTP 400 instead of 500.
• SemanticLinksManager invalidates the manager cache on create/delete so the UI sees fresh state immediately.

Test plan

• Unit: tests/unit/test_clean_truncated_turtle.py (10 cases)
• Integration: tests/integration/test_knowledge_routes.py covers truncated-Turtle recovery and unparseable rejection
• Component: linked-objects-panel.test.tsx (6), concept-relations-panel.test.tsx, concept-neighborhood-graph.test.tsx, business-terms.test.tsx (redirect)
• Manual: assign Data Product / Data Contract / Asset to a concept; verify navigation and removal
• Manual: open a concept with many neighbours; confirm the relations and linked-entities lists scroll past the cap and the neighbourhood graph remains readable in light + dark modes
• Manual: re-import an LLM ontology that previously truncated; confirm save succeeds and Settings → Rebuild Graph refreshes the tree

[mvkonchits-db]

Hi Lars, one of the CI tests failed plus 5 cents from my side (what to double check):

• settings_routes.py:364 — audit field rename user_email → requester_email. Confirm requester_email is the canonical field, else downstream audit consumers break silently.
• _get_asset_child_columns N+1 (one SELECT per hasColumn relationship per table-like asset).
• get_all_columns / search_columns paginate in memory rather than push to SQL.
• _resolve_asset_parents recurses without a visited-set.

[larsgeorge-db]

Thanks for the careful review @mvkonchits-db — circling back now that main has caught up. The branch is rebased onto current main (77e5c86), which dropped the duplicate data-catalog commit and inherited the recent fixes. Per-bullet response:

• CI: TypeScript Type Check (Database unused import) — Fixed on main by 375c906 (drop unused Database icon). Resolved by rebase; CI re-running now.
• settings_routes.py:364 audit field rename — Not actually a rename. The dict key is still 'user_email'; only the source attribute changed in e6cca03 because HandleRoleRequest only exposes requester_email. Downstream audit consumers continue to find details.user_email. PR feat(concepts): full-page browser/detail, links, neighbourhood, robust ontology save #340's actual change to settings_routes.py is the clear_demo_data SAVEPOINT-per-DELETE hardening, unrelated to that line.
• _get_asset_child_columns N+1 — Fixed on main by 9b92cb0: _get_columns_from_assets now eager-loads source_relationships.target_asset, and _get_asset_child_columns reads rel.target_asset directly. Inherited via rebase.
• get_all_columns / search_columns paginate in memory — Server-side offset/limit, faceted filters, and the removed 5000-record search cap landed in main via feat(data-catalog): server-side pagination, asset columns, hierarchy filters & expanded search #337 (0c318f8). The deeper "push everything to SQL" optimisation is acknowledged in feat(data-catalog): server-side pagination, asset columns, hierarchy filters & expanded search #337's review and tracked as [Perf]: Push data-catalog pagination/filters into SQL (perf follow-up to #337) #409.
• _resolve_asset_parents recurses without a visited-set — This maps to _collect_schemas_from_container in assets_manager.py (the only recursive descent on assets). Still un-guarded on main, and PR feat(concepts): full-page browser/detail, links, neighbourhood, robust ontology save #340 doesn't touch that file. Filed as [Perf]: _collect_schemas_from_container recurses without a visited-set #410 with a proposed visited-set + depth-guard fix patterned after _build_delete_tree in the same file.

PTAL when CI clears.