Emit session configuration event with loaded tools, instructions, and extensions at startup

## Describe the feature or problem you'd like to solve

The JSONL event stream lacks visibility into what the agent loads at startup — tools, instruction files, MCP connections, extensions, and skills are all resolved before the first turn but never reported in telemetry.

## Proposed solution

A new event (e.g. `session.config_loaded`) emitted once after initialisation, reporting what was loaded. This would allow anyone building observability or cost-analysis tooling on top of the event stream to answer questions that are currently impossible to answer:

- How many tool definitions are injected into the system prompt per turn, and what is their token cost?
- Which instruction files from `.github/instructions/` matched the current context, and what is their combined size?
- What fraction of loaded tools are actually invoked during a session?
- Which MCP servers connected and how many tools did each register?

Without this data, users optimising token budgets have no way to measure or reduce the overhead from tool/instruction loading — the largest controllable component of per-turn token consumption.

A suggested payload structure:

```jsonc
{
"type": "session.config_loaded",
"data": {
"tools": [
{ "name": "bash", "source": "builtin" },
{ "name": "github-list_issues", "source": "mcp:github" }
],
"toolCount": 84,
"instructions": [
{ "path": ".github/instructions/dev.instructions.md", "sizeBytes": 2048, "matchedPattern": "apps/**" }
],
"instructionCount": 2,
"instructionTotalBytes": 4096,
"extensions": [
{ "name": "my-extension", "location": "project" }
],
"extensionCount": 1,
"skills": [
{ "name": "docs-build", "source": "project" }
],
"skillCount": 5
}
}
```

Even a minimal version — just `toolCount` and `instructionCount` as integers on the existing `session.start` event — would be a meaningful improvement.

## Additional context

- The `tool.execution_start` event records each tool invocation (name + arguments) but not the catalogue of available tools, so "tools used" can be counted but "tools available" cannot.
- The `skill.invoked` event partially covers skills but does not address tools, instructions, or extensions.
- Organisations managing token budgets across multiple teams and repositories currently have no mechanism to identify which configurations cause excessive token consumption. This event would close that gap.
- This is a read-only telemetry addition — no changes to agent behaviour, just improved observability.