[Feature] Support progresssive disclosure when learn=true

### Describe the feature

When the response is so large that it gets written to a temporary file, tool discovery failures are more likely to occur. The model will use grep to understand the commands which may miss commands with unexpected names.

### Expected behavior

Consider progressively disclosing information about commands:

1. Namespace-level  `learn=true`  returns command names and short descriptions only.
2.  `learn=true`  with  `command=advisor_recommendation_summary` returns only that command’s full parameter schema.
3. (This may be fixed) Invalid commands return concise names and possibly nearest matches—never the complete catalog.

### Actual behavior

Large files.

### Reproduction Steps

https://github.com/microsoft/mcp/actions/runs/35278325710/job/105394183069. Evaluations with resiliencemanagement or monitor fail with errors similar to this.

### Environment

_No response_