Observability severity patch

Central severity_policy.py classifies explicit MCP failures, guardrail blocks, judge rejections, RAG failures and memory/cache errors.
Telemetry.event() accepts level and status_message and preserves WARNING/ERROR events as individual observations in compact mode.
Telemetry.span() handle supports set_status(level, message); failed MCP results mark their span ERROR without raising exceptions.
Long-term-memory read/write failures emit telemetry while preserving existing fallback behavior.
Guardrail blocked and negative judge evaluation emit WARNING when these helper methods are invoked.
Normal cache misses and empty RAG results are not automatically failures.

Scope / limitations

Existing call sites must invoke GuardrailTelemetry.blocked for guardrail blocking events; this patch does not rewrite every pipeline call site.
Redis cache currently has no injected Telemetry instance, so its existing logger errors are not yet Langfuse observations. A separate dependency-injection change is required to cover that path safely.
RAG database connection errors that are caught by external KBDB integrations without emitting a telemetry event cannot be classified automatically.
No live Langfuse ingestion test was possible in this environment. Verify both compact and verbose modes against your server.

PYTHONPATH=libs/agent_framework/src python -m pytest -q tests/test_observability_severity_policy.py