Validate installed MCP transport, optional torch integration, and backends #56

Parent epic: #52

Objective

Validate the built copick-mcp distribution through real FastMCP/stdin-stdout transport, exact optional torch integration, and the metadata-focused backend matrix.

Prerequisite

• Format-neutral metadata exploration parity is complete.

Work

• Build and inspect a wheel and source distribution from the lock.
• Install the wheel with exact copick==2.0.0-alpha.1 and copick-utils==2.0.0-alpha.1.
• Through an in-memory FastMCP client, list and call all 14 tools and validate content/schema serialization.
• Start the installed module over stdio and perform initialize/list/call/shutdown with protocol-only stdout and logs on stderr.
• Load/render help for all three setup entry points through the installed copick CLI using temporary configs.
• Install the torch extra with exact copick-torch==2.0.0-alpha.1 and verify installed torch command discovery/details.
• Verify get_nnunet_workflow_info() reports the optional package accurately without running a model, training, inference, or GPU callback.
• Run the full local metadata matrix.
• Run focused list/detail calls on S3-compatible, Secure Shell, and ML Croissant projects.
• Exercise repeated Secure Shell discovery across core's reconnect path without rebuilding the MCP server.
• Run a bounded public portal metadata smoke and optional Server Message Block coverage.

Acceptance criteria

• Clean installed base and exact torch-extra environments pass.
• All 14 MCP tools and three copick setup entry points load from the built wheel.
• Real stdio calls and failures contain only protocol frames on stdout.
• Exact alpha plugin discovery agrees with installed entry points and providers.
• Optional torch status/guidance is correct without command callback execution.
• Local, S3-compatible, Secure Shell, and ML Croissant metadata gates pass.
• The portal smoke is bounded and Server Message Block remains optional/non-gating.

Non-goals

• No volume download for backend/portal smokes.
• No processing, conversion, training, or inference execution.
• No GPU or model-artifact gate.