export harness: allowedTools selects the wrong configured tools, and model additionalParams are dropped

[amhegazy-mp4]

Description

agentcore export harness produces an agent that differs from the harness it exports in two ways:

1. allowedTools selects the wrong customer tools. In a harness, a bare pattern (shell, file_*) selects built-in tools, and @name selects a configured tool such as an MCP server or inline function. The export does the reverse for configured tools: allowedTools: ["@exa"] drops the exa MCP server, while the bare ["exa"] keeps it.
2. Model additionalParams are dropped. For bedrock, open_ai, and gemini models, export harness --arn omits the harness model's additionalParams with an export note, so the exported agent calls the model with different request parameters than the harness.

Steps to Reproduce

1. Create a harness with a remote MCP tool named exa and allowedTools: ["@exa"].
2. Run agentcore export harness --name <harness> (or --arn).
3. The generated main.py has no MCP client for exa.

For (2): export a deployed harness whose Bedrock model sets additionalParams; the export notes say they were omitted and model/load.py does not include them.

Expected Behavior

The exported agent gets the same tools the harness allows, and its model is called with the harness's additionalParams.

Actual Behavior

@exa drops the MCP server (and a bare exa keeps it); additionalParams are omitted.

Operating System

macOS