Add conversational Assistant panel to /query

[kosminus]

Summary

Adds a conversational Assistant panel on the /query page. Editors ask data questions and grow the semantic layer in plain language; the assistant only drafts — every write rides the existing REST + auto-embed endpoints, so authorization stays unchanged. MCP remains external-only.

One stateless POST /query/assistant turn runs a single structured-JSON LLM call that classifies intent and extracts a draft (no tool-calling loop, no provider changes, no chat-session table — history is passed from the client).

Intents

Intent	Action	Confirmed write
question	sql_preview (reuses generate_sql_only)	POST /query/execute-sql
glossary	glossary_draft	POST /connections/{id}/glossary
metric	metric_draft	POST /connections/{id}/metrics
dictionary	dictionary_draft	POST /columns/{column_id}/dictionary (per entry)
knowledge	knowledge_draft	POST /connections/{id}/knowledge
chat	—	plain message

All draft intents are editor-gated (ctx.has_role(ROLE_EDITOR)); viewers get questions only. Dictionary resolves (table_name, column_name) → column_id server-side against the schema cache and supports multiple value mappings per column, downgrading to a clear chat message when the column isn't found.

Design notes

• Mounted under /query (body connection_id, get_org_context) mirroring the existing query endpoints — inherits the /query/* rate limiter for free; the service's get_connection enforces read-authz.
• Agent returns a generic AssistantDecision{intent, message, payload}; per-intent normalizers validate/shape the payload and downgrade to chat when unusable.
• Frontend: AssistantPanel renders colour-coded, editable confirmation cards that POST to the existing create endpoints and invalidate the relevant query caches. QueryResultView extracted from QueryPage for reuse.

Test plan

• Backend: 131 passed (33 new assistant tests — agent normalizers per intent, downgrade-on-unusable-payload, service branching incl. dictionary column resolution + viewer-blocked-before-resolution). ruff clean.
• Frontend: tsc --noEmit clean, ESLint clean, npm run build succeeds.
• Manual: try "What is total ECL by stage?", "Add a glossary term: NPL means loans where stage = 3", "Define a metric: ECL coverage ratio = sum(ecl)/sum(exposure)", "In the stage column, 1=performing, 2=underperforming, 3=non-performing", "Remember this about IFRS 9: …".

Plan / as-built notes in docs/assistant-plan.md.

🤖 Generated with Claude Code