fix(provider-codex): sanitize Responses message ids before dispatch

[M1k0t0]

Summary

• strip non-OpenAI message ids from Codex Responses input before upstream dispatch
• preserve valid msg* message ids and leave non-message item ids untouched
• handle both typed message items and typeless easy input messages when they reach the Codex provider boundary

Why

Codex can echo local/review rollout message ids that the upstream Responses API rejects. Those ids are client-local, not upstream-owned, so Floway should remove them before sending the request to Codex upstream while preserving the message content.
This PR is scoped to the Codex provider wire boundary. A separate gateway-side normalization fix is still needed for stored native Responses requests where typeless { role, content } messages are staged before provider interceptors run.

Reproduction

1. Run /review and wait for the review to complete.
2. Send another message in the same Codex conversation.

[Menci]

But, how does it work on Codex official client+server?

[M1k0t0]

I dug into this more. The relevant official Codex behavior has two paths.

For normal Codex Responses requests, the client strips item ids when the item_ids feature is off, and that feature is disabled by default. Therefore the official Codex client does not normally send local rollout ids such as review_rollout_user.

However, when Codex thinks it is connected to an Azure Responses endpoint, for example because the base URL contains azure-api., it sets store: true, and the id-stripping path is skipped when store is true. After that, the official Codex CLI really does send the local /review rollout ids to Floway even though features.item_ids is not enabled.

Floway then forwards to the Codex upstream, where those local ids are not valid Codex message ids. This patch handles that compatibility boundary: preserve upstream-shaped message ids (msg...) but drop client-local message ids like review_rollout_user before forwarding to Codex.

This PR is still draft for a separate reason: for stored native Responses requests, Floway’s gateway can currently fail while staging typeless { role, content } message inputs before the request reaches the Codex provider boundary. I'd like to fix that upper gateway normalization path first, then come back and finalize this provider-level sanitizer change.

Refs:

• Guard that skips id stripping when item_ids_enabled || store:
  codex-rs/core/src/client.rs#L871-L878

• store is set from provider.is_azure_responses_endpoint():
  codex-rs/core/src/client.rs#L852-L860

• azure-api. is one of the Azure URL markers:
  codex-rs/codex-api/src/provider.rs#L116-L126