## What happened?

When calling `litellm.completion()` (or the proxy `/v1/chat/completions`) for an OpenAI model with the OpenAI Chat Completions parameters `store` and/or `prompt_cache_key`, LiteLLM accepts the request without any error or warning, but **neither parameter is included in the outbound HTTP request to the provider**. `safety_identifier`, passed the same way, IS forwarded correctly.

This is silent data loss of documented OpenAI parameters:

- `prompt_cache_key` — OpenAI's cache-routing hint ([prompt caching guide](https://platform.openai.com/docs/guides/prompt-caching)). Losing it degrades cache hit rates for high-volume workloads without any visible symptom. - `store` — explicit opt-out/in for OpenAI request storage. Losing `store: false` silently discards a data-governance declaration.

Both params were added to the *supported params* lists in PR #20989 (so `drop_params`/`UnsupportedParamsError` no longer reject them), but the forwarding plumbing was never completed — see root cause below. Verified on `v1.90.2` and still present on `main` as of 2026-07-14.

## Minimal reproduction (no proxy needed, no API key needed)

```python import json, threading from http.serv…