Add Ask AI tab (backend OpenAI SDK, per-table summary memory, reuses graph-data job/poll mechanism)

[YanivZalach]

Blocked: this can only be started after #100 is done.

Summary

Add an "Ask AI" tab to IceGraph: a chat-style page, alongside Graph/Metadata/Timeline/FileTree, where a user can ask free-form questions about the currently-loaded table and get a full AI-generated answer back.

Confirmed requirements

There will be a custom prompt in the backend that will be on top of the user prompt.

• Backend adds the official openai Python SDK. The AI endpoint, model, and API key are configured entirely through environment variables (backend/constants.py-style config, same pattern as SPARK_REMOTE). No provider picker, no endpoint/token field anywhere in the UI.
• Frontend gets a new page: an AI chat assistant.
• Each question automatically includes context about the currently-loaded table (schema, partition spec, node/file counts).
• As the user chats, the conversation has memory across turns, but what's persisted to localStorage is a running summary, not the raw message list. Every request sends the question plus the current summary; every response returns both the answer and an updated summary, which overwrites what's in localStorage.
• localStorage is scoped per table (keyed by table name + snapshot range, same key style as the existing graphData_<table>_<start>_<end> cache key).
• The relay reuses the existing async job/poll mechanism (jobs dict + ThreadPoolExecutor + per-job token) that /api/v1/graph-data already uses. A synchronous call isn't acceptable because a slow AI response would block the whole app for every user.
• Responses are returned in full — no streaming.

Backend

New dependency: openai (official Python SDK). Its base_url override also covers Azure OpenAI and OpenAI-compatible self-hosted servers (Ollama, vLLM), so one client code path is enough. Add via uv add openai in backend/.

backend/constants.py — new env-configurable defaults, same pattern as MAX_SNAPSHOTS_TO_COMPUTE:

Variable	Default	Purpose
AI_API_BASE_URL	""	Base URL passed to the openai client. Empty = feature disabled.
AI_API_KEY	""	API key passed to the openai client. Empty = feature disabled.
AI_MODEL	"gpt-4o-mini"	Model name to request.
AI_REQUEST_TIMEOUT_SECONDS	60	Timeout passed to the openai client.

No new token-header constant needed — reuse the existing JOB_TOKEN_FIELD (X-IceGraph-Job-Token) for ask-ai jobs too.

Frontend

• A new "Ask AI" tab, in the same nav position as Graph/Metadata/Timeline/FileTree, reachable by keyboard shortcut 5.
• A chat interface: the user types a question, submits it, sees a loading state, then the full AI answer appears in the conversation.
• Each question automatically carries the currently-loaded table's context (schema, partition spec, node/file counts) — the user never has to supply this themselves.
• Conversation memory persists per table (scoped by table + snapshot range) across page reloads, as a running summary rather than the raw message transcript.
• If the AI backend is unavailable or misconfigured, the tab shows an inline error without breaking the rest of the app.
• Visually and structurally consistent with the existing tabs (same nav/typography/error-state conventions already used elsewhere in the app).

Per CLAUDE.md's standing convention for this repo: DocsPage.jsx gets a section for the new tab, and claude-plugin/skills/icegraph/SKILL.md gets the new /table/ask-ai route.

[Razberrry]

I'd actually recommend using some kind of library for this like assistant-ui. Let them worry about all of this it's plug and play