fix(codex): Azure Foundry rejects replayed reasoning from more than one response; classify and prune

[erosika]

Summary

On Azure Foundry, gpt-6-astra returns HTTP 400 "Conflicting authenticated continuation identities" on the first request of every turn after the first, and the session never recovers. Fixes #105369.

What the endpoint rejects

That request replays the encrypted reasoning item of every earlier response. I replayed the persisted items of a failed session against the resource-level endpoint. Any one item alone returns 200. Any two together return 400. Restoring the stripped rs_ ids changes nothing, so #67667 does not cover this case. In-turn calls never hit it: _is_post_tool_replay already suppresses replay on Azure when the request ends on a tool result. The background review named in the issue is coincident; one failed session had no review between its turns.

Transport

On Azure Foundry, build_kwargs keeps codex_reasoning_items only on the newest assistant row that has any. include still requests reasoning.encrypted_content, compaction checkpoints stay on every row, and the canonical messages are not mutated. Other Responses endpoints are unchanged. The gate is its own predicate, provider azure-foundry or either Azure host, so narrowing _is_azure_foundry_responses in #101243 does not switch it off.

Classifier

The message now maps to invalid_encrypted_content. The existing recovery in turn_recovery strips the replay state and retries once, instead of aborting as a non-retryable client error.

Validation

I rebuilt the failing session from state.db through the transport and sent both payloads live. Unpatched: 5 reasoning items, 400. Patched: 1 item, 200. The transport, classifier and codex request suites pass, 326 tests.

Important

Verified against *.openai.azure.com/openai/v1 only. #101243 targets a different 400, invalid_payload, on the project-scoped gateway. A one-item control passes on both endpoints, so it cannot show that an endpoint tolerates multi-item replay.

Follow-up

auxiliary_client.py converts messages for Responses without this pruning. This PR does not touch it.