Phase 2 (Milestone 1): Durable analytics artifacts

[kosminus]

Summary

Implements Milestone 1 of Phase 2 from planfull.md — turning one-shot answers into durable, owned, re-runnable objects.

• Saved queries — named, owned NL question + pinned SQL with typed parameters ({{date_from}}, {{region}}), versioned, clone/fork.
• Result snapshots + cache — results persisted in a Postgres table that doubles as a cache (keyed by sha256(final_sql + params + connection_id), TTL RESULT_CACHE_TTL_SECONDS default 300 + manual refresh), so re-runs don't re-hit the warehouse.
• Charts — a persisted chart config per saved query, rendered with Recharts (table/line/bar/area/pie/scatter).
• Export — client-side CSV/JSON on results; backend CSV/JSON/XLSX for saved queries.

Dashboards (Milestone 2) are intentionally out of scope here.

Backend

• Models SavedQuery, Chart, ResultSnapshot — connection-scoped, organization_id-keyed, matching the Phase-1 metadata pattern (workspace isolation flows through the connection via require_connection_read/write).
• Migration 005_durable_artifacts — 3 tables + (sql_hash, taken_at DESC) cache index.
• saved_query_service.render_sql — type-safe parameter substitution (numbers/booleans inlined, strings/dates escaped + quoted); the rendered SQL still passes through check_sql_safety. Cache-first run_saved_query reuses query_service.execute_raw_sql.
• Endpoints: CRUD + /run, /clone, /export?format=csv|json|xlsx, and a charts sub-resource. openpyxl added as the optional [export] extra.

Frontend (Recharts added)

• SavedQueriesPage (list + Run/Edit/Clone/Delete), run drawer (param form + cache badge + chart panel), form modal with a params editor, ChartView, SaveQueryModal.
• "Save query" + client-side CSV/JSON export on the Query result view; nav item + /saved-queries route.

Notable decisions

• Result cache = Postgres table (keeps the default self-hosted deploy Redis-free).
• Charts are managed inside the saved-query view, not a separate top-level page (a chart without a query is meaningless).
• Artifacts are connection-scoped with organization_id rather than a direct workspace_id — matches the actual Phase-1 implementation rather than the idealized schema in planfull.md §5.

Verification

• Backend: full pytest suite passes (incl. 15 new render_sql/sql_hash tests); ruff + format clean; mypy clean on new files.
• Live (Docker + real Postgres + IFRS 9 sample DB): migration 004→005 ran on startup; create → run fresh → cached → refresh-bypass → distinct cache key per param verified; chart create, CSV/JSON/XLSX export, and clone all work.
• Security: a '; DROP TABLE counterparties; -- parameter value was neutralized (escaped to a literal and caught by the static blocklist); the table survived with 20 rows. Bad-type params → 422.
• Frontend: ESLint clean; tsc -b && vite build succeeds.

🤖 Generated with Claude Code