ErikAI: analyze_server should return complete analysis, not breadcrumb trail

[erikdarlingdata]

Problem

analyze_server currently returns findings with next_tools — a list of MCP tools the AI client should call for deeper investigation. This is backwards. The engine has access to all the same data. It should do the drill-down itself and return a complete analysis.

Example: the engine detects a CPU spike and returns:
```json
{
"story_path": "CPU_SPIKE -> SOS_SCHEDULER_YIELD",
"next_tools": ["get_cpu_utilization", "get_top_queries_by_cpu"]
}
```

The AI client then has to make 2 more tool calls, interpret the results, and correlate them. The engine already knows WHERE to look — it should look there and include the results.

Proposed Behavior

analyze_server should:

1. Run the inference engine (collect → score → traverse) as today
2. For each finding, follow its own next_tools recommendations internally
3. Include the drill-down results in the finding output — e.g., the top queries during a CPU spike, the deadlock graph for a deadlock finding, the blocking chain details for lock contention
4. Return a self-contained analysis that an AI client can narrate directly

Benefits

• Single tool call produces a complete diagnostic picture
• AI client doesn't need to understand which follow-up tools to call
• Reduces round-trips from ~5-10 tool calls to 1
• The engine's expertise is in the drill-down logic, not just the symptom detection

Constraints

• Keep next_tools in output for clients that want to do their own investigation
• Response size — deep drill-down data could be large; may need a detail_level parameter (summary vs full)
• Individual tools (get_deadlocks, etc.) remain available for manual investigation

Relates to

• Lite MCP coverage gaps (#TBD) — the engine can only drill down into data that has tool/query support
• The query_snapshots gap is the most acute example: CPU spike detected but no way to find the causal query