### Check for existing issues

Searched `litellm/litellm` issues for existing reports on the Headroom compression guide — found none specific to this.

### What happened?

The Headroom guardrail deploy guide (`docs.litellm.ai/docs/proxy/headroom#deploy-headroom`) pins `headroom-ai[proxy]==0.27.0` in its example Dockerfile. Following the guide exactly as written has two problems:

1. `headroom-ai==0.27.0` predates a crash fix in Headroom's own repo (`headroomlabs-ai/headroom#1114`, fixed in `0.31.0`/`0.32.0`) for its `HeadroomCallback` integration path — anyone following the pinned version can still hit `AttributeError: 'HeadroomCallback' object has no attribute 'async_post_call_success_hook'` on real requests, even though that issue is already closed upstream as fixed.

2. More importantly: following the guide's own `guardrail: headroom` / `mode: pre_call` example exactly as documented — correct network topology (sidecar), correct request schema once the undocumented required `model` field is added — results in `tokens_saved: 0` for realistic tool-output/RAG/JSON-shaped payloads, which is presumably the traffic this guardrail is meant for. Root cause is two independently-gated defa…