### What happened?

On a self-hosted proxy (`v1.87.0`), a **custom `pre_call` guardrail** (a `CustomGuardrail` subclass that masks credentials/PII in the request and writes the masked body back into `data`) runs and masks correctly on **both** endpoints:

- OpenAI `/v1/chat/completions` (`call_type: acompletion`) - native Anthropic `/v1/messages` (`call_type: anthropic_messages`)

However, the **guardrail bookkeeping is only persisted for the OpenAI route**. On the native `anthropic_messages` route, `guardrail_information` is `NULL` on **successful** requests, so the Admin UI → session → **Logs** tab shows no guardrail information for those requests, even though the guardrail demonstrably executed (the input was masked upstream).

This makes guardrail execution effectively **invisible to auditing** for the native `/v1/messages` route.

### Evidence

Comparing two `LiteLLM_SpendLogs` rows for the same guardrail config, one per route (both rows carry the **same set of metadata keys** — the difference is in the value):

| field | `/v1/messages` (`anthropic_messages`) | `/v1/chat/completions` (`acompletion`) | |---|---|---| | `metadata.guardrail_information` | **`NULL`** | populated (a…