Guard ignored runtime configs with schema versions

[XuweiDing04]

Problem

config.yaml is intentionally ignored, so git pull updates config.example.yaml but leaves an existing user runtime config untouched. After #112, a stale local file can still omit allowed_openai_params: [reasoning_effort] for gpt-5.6-sol, causing LiteLLM to reject reasoning_effort with a 400 even though the repository itself is current.

Solution

• Add a tracked top-level config_schema_version to config.example.yaml.
• Validate root and explicit --agent-config files before run, run-eval, or eval uses them.
• Fail fast on missing, stale, newer, invalid, or non-mapping configs with safe migration instructions.
• Keep --help usable even when the local runtime config is stale or malformed.
• Document the migration rule: back up local settings, merge the latest template, then update the version. The code never overwrites credentials or user customization.

Verification

• uv run pytest -q — 845 passed, 2 skipped.
• Ruff F/import checks and git diff --check passed.
• The reported LB2-0107 command succeeds with --dry-run and resolves the expected task.
• Real browser-use + gpt-5.6-sol + lexmount smoke reached multiple model/browser steps without the previous LiteLLM 400; the long benchmark was then stopped intentionally and its Lexmount session was deleted.