Summary

MCP tool results from stackql mcp currently carry the typed payload in
structuredContent and a markdown rendering in the text content block
(pkg/mcp_server/render). That is spec-correct, but a large class of MCP
clients - LLM hosts, scripting clients, and SDK helpers - only consume the
text content. For those clients the markdown table/KV rendering is lossy
and unparseable, and there is no way to opt into machine-readable text.

Current behavior

tools/call run_select_query returns:

• structuredContent: {"rows": [...]} (typed DTO)
• content[0].text: a markdown table (RenderTable) or KV records
  (RenderKV), e.g. # Pull Result\n\n## Record 1\n...

Proposed enhancement

Either (or both) of:

1. A per-call argument on the query/metadata tools, e.g.
   format: "markdown" | "json" (default markdown, preserving current
   behavior), where json renders the same DTO as compact JSON into the
   text content block:

   {
     "name": "run_select_query",
     "arguments": {"sql": "SELECT 1", "format": "json"}
   }

2. A server-level render default in --mcp.config, consistent with the
   existing config surface:

   {"server": {"render": "json", "audit": {"disabled": true}}}

The per-call argument is the more useful of the two for mixed workloads
(human-facing hosts want markdown, embedded/scripting clients want JSON in
the same session).

Context

Encountered while building the embedded Rust client (stackql-mcp-rs):
clients that follow the MCP spec can and should prefer structuredContent
(that is what we ship), but interop with text-only consumers currently
requires parsing markdown. Observed on stackql v0.10.500, MCP stdio mode.