Production refinement: backend correctness, dashboard polish, build gates

[askmy-stack]

Summary

End-to-end refinement following the orientation in CLAUDE.md / DECISIONS.md / MISTAKES.md. Backend bugs fixed, dashboard UX elevated, build pipeline tightened. No public API changes, no version bumps, make demo semantics preserved.

Full audit + checklist: docs/REFINEMENT_AUDIT.md.
New decision logged: D-017 (causal-chain depth literal + tsc build gate).

Backend (correctness + perf)

• graph/query.py — trace_causal_chain rendered with a clamped depth literal (1..8). Neo4j 5 rejects [:SUPERSEDES*1..$param]; the previous query would have raised at runtime for every /decisions/{id}/chain call against a real graph.
• api/main.py::/health and api/contradictions.py now reuse the shared async driver via MemoryService instead of opening a new neo4j.GraphDatabase per request. Contradictions queue is now async def.
• pipeline/extraction_worker._processed switched from an unbounded set[str] to a bounded LRU (_BoundedSeenCache, capacity CORTEX_DEDUP_CACHE_SIZE, default 262144). Prevents the memory leak that would have surfaced after long worker uptime.
• Dead imports removed (memory/semantic.py, intelligence/contradiction_detector.py, api/decisions.py). Touched files are ruff-clean.

Frontend (UX + perf + build gates)

• package.json::build now runs tsc --noEmit && vite build; tsconfig.json enables noUnusedLocals/noUnusedParameters. Previously TS errors silently shipped because vite build only strips types.
• App.tsx code-splits ExploreView, AgentsView, ReviewView via React.lazy + <Suspense>. Four-chunk build, total ~165 kB / 53 kB gz.
• New shared primitives — Skeleton, StateView, TypingIndicator (no new deps).
• Skeletons + accessible empty/error states across all data-fetching views; ReviewView auto-loads on mount; AskView debounces its character counter and memoises the result list.
• DecisionCard chevron rotates 180° via motion token; component is React.memo-wrapped.
• AssistantPanel typing indicator with aria-live, cleanup on unmount + on rapid re-send, Send disabled when empty.
• MemoryGraph nodes now activate on Space (not just Enter) and expose aria-pressed/aria-label.
• Design system polish: bumped --text-muted above WCAG AA, added --focus-ring + global :focus-visible, shimmer + prefers-reduced-motion overrides, responsive at 360 / 768 / 1100 / 1440 with proper mobile sidebar (44 px tap targets, sticky scrollable tabs).

Hard constraints honored

• No security weakened (signature verification + RBAC unchanged; contradictions still RBAC-filtered inside the graph layer).
• No tests removed; baseline tests updated for the new async contradictions path. 228 → 235 pytest passing, ~80% → 81.07% coverage.
• No secrets, no major version bumps, no heavy deps added.
• make demo / scripts/demo.sh semantics untouched. Frontend Docker image still serves on port 3000.

Test plan

• uv run pytest tests/ -q — 235 passed, 81.07% coverage (coverage gate ≥70%).
• cd frontend && npm run build — clean (tsc --noEmit → vite build, 52 modules, four chunks).
• uv run ruff check on touched modules — no warnings.
• Reviewer: make demo + manual click-through on Home / Ask / Memory map / For agents / Review + mobile breakpoint at 360 / 768.
• Reviewer: confirm /decisions/{id}/chain against a live Neo4j now returns nodes rather than the silent driver error.

Intentionally deferred

• Async TimescaleDB driver for memory/episodic.append_raw_event (worker still uses sync psycopg; out of scope for a UX-focused pass).
• Frontend test framework (Vitest + Testing Library) — Cortex's stated test home is /tests/ (Python). Recommended as a follow-up PR.
• Storybook / visual regression — would violate the "no heavy new deps" constraint.

See docs/REFINEMENT_AUDIT.md for the full audit grid.