feat(reins): knowledge base isolation — every base answers only for itself (CLEAN-48)

[maksymhryzodub-prog]

Summary

Implements CLEAN-48 — the knowledge-workspaces research outcome: a knowledge base becomes a real isolated area instead of a label on a shared pool, and the module stops needing a briefing.

• Isolation (US1): one LightRAG instance per base, provisioned through the same Argo path as agent pods (reins/instance sub-slice mirroring workflow; image pinned by digest; slot-sized requests). Query, graph and entity list are base-scoped; a migrated base never falls back to the shared pool — a down instance is a stated 503, not silently borrowed content. Empty/no-coverage bases return answer: null, reason: no_relevant_content instead of a generated answer. References resolve to Source rows (file_source carries the source id). The agent tool lists only bound bases, refuses unbound ids, attributes every block, and names unreachable bases. SC-001/SC-002 are executable Jest specs.
• Transition: one-time resumable per-base re-index (reins/migration) rebuilt entirely from Ranch's own storage — the operator supplies nothing. Reads stay on the shared pool until a base flips to done; the shared deployment remains the rollback until decommission (deliberately deferred, gated on every base migrating).
• Finding bases (US2): server-side searchable, paged list with per-base context (sources, size, state); the agent screen edits bindings in place, marks template inheritance, and surfaces bindings to deleted bases instead of silently dropping them.
• Clarity (US3): entityTypes/relationshipTypes removed from schema and contract (never sent anywhere in the product's history); retrieval tuning folded behind a disclosure with trade-off copy; the base header shows what it holds and whether it can answer.
• UI defects (US4/US5): active tab fixed via NuxtLink custom slot (root cause: two utilities setting the same property at equal specificity); /knowledges/:id lands on Sources; the entity picker is a virtualized reka-ui combobox over the base-scoped labels endpoint with server-side search.
• Honest status (US6): indexStatus derived from per-source states (empty/indexing/partial/ready); the misleading indexed boolean removed; per-source failure reasons with a single-source retry endpoint.

Spec, plan, research, contracts, task list and a verification log live in specs/007-knowledge-workspaces-research/.

Test plan

• 182 Jest tests green (21 suites): manifest builder, routing policy, isolation integration (SC-001), adversarial tool set (SC-002), state mapping, readiness rollup
• tsc --noEmit clean; OpenAPI regenerated into both consoles at every breaking step
• Live HTTP run against the local stack (docker LightRAG + Ollama, mock provider): create → provision → ingest → queued→processing→indexed via track_status → query answers with resolved reference; empty base refuses without touching retrieval; scoped label search; reindex 202; derived rollup
• emptyDir sufficiency proven live (both working dirs 0 bytes through ingest; query answers after restart with empty dirs)
• Cluster run of quickstart scenarios 1–3 and 8 with the Argo instance provider (local mock cannot reproduce isolation by construction — stated in quickstart.md)
• Manual browser click-through of the admin screens (SSR compiles all routes; API behavior behind each screen verified live)
• T045 decommission of the shared pool — operational step after every base reaches migrationState: done on the live installation

🤖 Generated with Claude Code