fix(agents): keep Codex reasoning payloads stateless

[kevin-dp]

Summary

#4557 forced store: true on built-in reasoning payloads for both openai and openai-codex so OpenAI Responses continuations can replay rs_* reasoning items. But the ChatGPT-login Codex endpoint is stateless-only and hard-rejects any stateful request:

pi-adapter message_end ERROR stopReason=error errorMessage={"detail":"Store must be set to false"}

Since #4557 merged, every Codex (gpt-5.x via ChatGPT login) run fails immediately, even a plain "Hi" with no tool calls.

Fix

Scope the stateful default to the regular OpenAI Responses API only — the openai-codex payload keeps the upstream store: false default (pre-#4557 behavior). This means the rs_* not found continuation issue #4557 addressed remains possible on Codex, but stateful requests can never work against that endpoint anyway; a stateless fix there would need the encrypted reasoning-content replay path (explicitly a non-goal in #4557).

Test plan

• pnpm exec vitest run test/model-catalog.test.ts in packages/agents (15 passed) — adds a regression test that openai-codex payloads get no store override (and an incoming store: false survives), while the existing openai store-true coverage still passes
• pnpm typecheck clean in packages/agents
• Manually reproduced against the desktop app: gpt-5.5 runs failed with the store error before this change

🤖 Generated with Claude Code