Spike: scope natural language querying

## Goal

Investigate and scope the natural language querying feature before any implementation begins. Questions to resolve:

- **UI placement** — where does the NL query experience live? Search sidebar, query editor mode, or a dedicated surface?
- **Interaction model** — how does it function? One-shot NL → query translation the user can inspect/edit, a conversational flow, or direct NL → results?
- **Deployment** — how does a user deploy this? Graph Explorer is a client-side app with a thin proxy; does inference go through the proxy server, a sidecar, or directly from the browser?
- **Bring-your-own backend** — can users supply their own model endpoint and credentials (Bedrock, Anthropic, OpenAI-compatible, local)? What's the configuration and credential-storage story given all persistence is client-side?
- **MCP story** — is there an MCP server exposing schema/connection/query tools so external agents can drive Graph Explorer, and/or does Graph Explorer act as an MCP client?
- **Context strategy** — what schema/sample context does the model need to generate correct Gremlin/openCypher/SPARQL, and how does it fit within existing schema sync?

## Related Issues

- Part of #690
- Related to #810 — query editor enhancements; a likely UI surface
- Related to #1596 — database capability detection; NL generation needs to know dialect/version constraints

> [!IMPORTANT]
> Internal only — this issue is maintained by the core team and is not accepting external contributions.

Start by reading the questions in this issue and the related issues #690, #810, and #1596. Investigate the UI, deployment, backend configuration, MCP, and schema-context options, then document decisions and recommendations that define the feature scope before implementation begins.

Written by the indexing model from the issue text.