fix(codex): cross-turn reasoning cache missing for custom_tool_call and tool_search_call

[thedavidweng]

Problem

The cross-turn reasoning cache in Codex Chat history only handles function_call type items. When Codex uses custom_tool_call (e.g. apply_patch) or tool_search_call, their reasoning_content is not cached and cannot be restored in follow-up turns.

This causes upstream providers like Kimi/Moonshot and DeepSeek to reject requests with errors like reasoning_content is missing in assistant tool call message, because the proxy cannot reconstruct the reasoning context that was present in the original tool call.

Current Behavior

In src-tauri/src/proxy/providers/codex_chat_history.rs:

• cached_function_call() (line ~439) only matches type == "function_call" and returns None for all other types
• record_response() (line ~57) uses .filter_map(cached_function_call), so only function_call items are recorded
• enrich_request() (line ~110) only matches "function_call" and "function_call_output", skipping "custom_tool_call", "tool_search_call", and their *_output counterparts

Expected Behavior

The reasoning cache should generalize to the full tool-call triad:

• function_call / function_call_output (already handled)
• custom_tool_call / custom_tool_call_output (missing)
• tool_search_call / tool_search_output (missing)

This ensures that when Codex sends a follow-up request with previous_response_id plus a custom_tool_call_output, the missing custom_tool_call with its reasoning_content is restored from the cache, matching the existing function_call behavior.

References

The streaming layer (streaming_codex_chat.rs) and transform layer (transform_codex_chat.rs) already handle all three tool types correctly. The gap is specifically in the cross-turn history cache.