[Bug]: Codex Luna Reserve fallback is missing from orchestrator v2

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Check out PR #2829.
2. Use a Codex account with ordinary quota exhausted and Luna Reserve available.
3. Start a Codex thread through orchestrator v2.
4. Select gpt-5.6-luna with Medium reasoning.
5. Send a prompt.

Expected behavior

Orchestrator v2 should match the official Codex CLI. It should detect Luna Reserve eligibility and send the turn using hidden wire model gpt-reserve.

The visible model and reasoning effort should remain gpt-5.6-luna and Medium. T3 should return to the ordinary model after Codex reports that ordinary usage is available again.

Actual behavior

#11253 tracks the same missing behavior in the current orchestrator.

I checked PR #2829 at 89459e384ad7ffa3b9dda613e9cfebf32a8979e1. The new orchestrator still uses the generated Codex protocol from before Luna Reserve support.

Required behavior	PR #2829
Send supportsLunaReserve: true	Missing
Decode ordinaryUsageAllowed	Missing
Decode the luna_reserve banner	Missing
Load hidden model gpt-reserve	Missing
Route eligible turns to gpt-reserve	Missing
Restore the normal model after recovery	Missing

CodexAdapterV2 copies the selected model directly into turn/start. It listens for ordinary rate-limit notifications, but never performs a Reserve-capable usage read or switches models.

Failed quota-exhausted turns are terminalized instead of triggering a Reserve refresh.

Impact

Blocks work completely

Version or commit

PR #2829 at 89459e384ad7ffa3b9dda613e9cfebf32a8979e1

Environment

Compared against openai/codex main at bc5957eac9e89e66f990ed490d11e625a4a3b02c. The shared generated-protocol work can be implemented with #11253. Orchestrator v2 still needs its own session state, turn routing, queued-turn handling, and recovery tests.

Logs or stack traces

Screenshots, recordings, or supporting files

No response

Workaround

No response

Start in apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts, especially the turn/start handling around lines 696-709, and compare the behavior with openai/codex main. Trace the usage read, session state, turn routing, queued turns, and recovery paths. Done means eligible Luna Reserve turns use hidden model gpt-reserve while preserving the visible model and reasoning, then return to the ordinary model after recovery, with recovery tests added.