Add Piston sandbox plugin

[eric-tramel]

What

Adds data-designer-sandbox-piston, a self-contained Data Designer plugin that registers the code-sandbox column type. The column executes source code from an existing dataset column through a configured local or remote Piston API and returns structured execution metadata including stdout, stderr, output, exit code, status, timing, and memory fields.

The package also includes a stdio MCP server exposing a run_code tool, helper config for creating Data Designer MCP providers/tool config, documentation, and Docker-based local/remote deployment examples.

Why

This makes the Big Iron sandbox/code-execution pattern available in the public plugin repo without Big Iron-specific deployment assumptions. Users can point Data Designer at a Piston endpoint running locally on macOS/Linux or at a separately managed remote service.

Usage

Column execution:

builder.add_column(
    name="sandbox_result",
    column_type="code-sandbox",
    target_column="python_code",
    language="python",
    version="3.12.0",
    sandbox_url="http://localhost:2000",
)

MCP helper:

from data_designer_sandbox_piston import SandboxMCPConfig

sandbox_mcp = SandboxMCPConfig(
    name="sandbox",
    sandbox_url="http://localhost:2000",
    language="python",
    version="3.12.0",
    result_fields=["stdout", "stderr", "exit_code"],
)

mcp_provider = sandbox_mcp.to_provider()
tool_config = sandbox_mcp.to_tool_config()

Local Piston helper:

plugins/data-designer-sandbox-piston/scripts/run-local-piston.sh
curl -X POST http://localhost:2000/api/v2/packages \
  -H 'Content-Type: application/json' \
  -d '{"language":"python","version":"3.12.0"}'

How

The implementation uses a shared aiohttp Piston client, structured SandboxOutput parsing, an adaptive concurrency controller with bounded retry/backpressure behavior, and a synchronous Data Designer column generator wrapper that preserves DataFrame row ordering. The MCP server launches as a stdio process and reads runtime settings from environment variables supplied by SandboxMCPConfig or the caller's environment.

The local Docker helper now mounts /piston to a Docker volume so stock Piston has a writable package store. Runtime defaults align with stock Piston's 3000 ms run limit, and python_packages is treated as declarative metadata for prebuilt custom Python runtimes rather than an implicit runtime-build request.

This PR intentionally leaves catalog/plugins.json unchanged; this repo's scaffold/release validation treats catalog registration as a first-release step.

Validation

• data-designer config list confirmed the gpt-5.5 model alias.
• Started local Piston via plugins/data-designer-sandbox-piston/scripts/run-local-piston.sh and installed python 3.12.0 with Piston's package API.
• Direct Piston execute smoke test returned stdout 55 with exit code 0.
• Data Designer column-mode workflow used gpt-5.5 to generate Python code, then code-sandbox executed it in Piston and returned stdout 650 with exit code 0.
• Data Designer MCP-mode workflow used gpt-5.5 with tool_alias="sandbox"; the trace shows one run_code call and returned stdout 6469693118 with exit code 0.
• make format
• make plugin-docs
• make test-plugin PLUGIN=data-designer-sandbox-piston
• make validate
• make check
• make lint
• make test-devtools
• make test-plugins
• make docs