fix(gateway): restore MoA one-shot model override on failed turns

[srojk34]

Summary

• When a MoA one-shot turn (`/moa `) fails with an exception in `_handle_message_with_agent`, the MoA model override is never restored — all subsequent messages silently route through MoA reference fan-out permanently
• The restore data is stored on the per-turn event object (`event._moa_restore_override`), which is discarded after the turn — unlike CLI/TUI where the restore flag persists on `self`/`session` and fires on the next successful turn

Root cause

The MoA one-shot restore ran inside the `try` block (line ~9032), after `_handle_message_with_agent` returns. When that call raises an exception (agent init failure, interpreter shutdown, OOM), execution jumps to the `finally` block which only calls `_release_running_agent_state` — the restore is skipped.

Since `_moa_restore_override` lives on the event object (not on the runner or session), it is permanently lost when the event goes out of scope. The next message creates a new event with no restore data, so the MoA override persists silently.

Impact

• Every subsequent gateway message fans out to reference models + aggregator (higher cost, higher latency)
• No user-visible indication that the model switched permanently
• Affects all gateway platforms (Telegram, Discord, Slack, etc.)
• Recovery: manual `/model ` switch

Changes

• `gateway/run.py`: move the MoA one-shot restore from the `try` block to the `finally` block so it fires on every exit path (success, exception, interrupt)
• `tests/gateway/test_moa_one_shot_restore.py`: 4 tests covering restore-on-success, restore-on-exception, restore-none-clears-override, and no-restore-for-non-moa-turns

Test plan

• `test_restore_fires_on_success` — successful turn restores previous model
• `test_restore_fires_on_exception` — failed turn (RuntimeError) still restores
• `test_restore_none_clears_override` — user with no prior override gets override removed
• `test_no_restore_when_not_one_shot` — normal turns don't touch model overrides

[tonydwb]

Code Review Summary

Verdict: Approved

Moves the MoA one-shot model restore logic from the success path into the finally block so a failed turn (exception) still reverts the model override. Complements PR #53548 (which makes /moa one-shot only).

Looks Good

• The restore logic in finally is idempotent (pop-on-absent is harmless)
• Placed before _release_running_agent_state to ensure cleanup ordering
• Comprehensive test suite: tests for success, exception, None restore, and non-MoA turns
• The 4 test cases in test_moa_one_shot_restore.py cover the key scenarios clearly

---

Reviewed by Hermes Agent

[teknium1]

Merged via #53559 — your commit was cherry-picked onto current main with your authorship preserved (2f29e3cfc). We added a small follow-up on top (4e0788783): extracted the restore into a _restore_moa_one_shot() helper so it's unit-testable without re-implementing it, restored the load-bearing #28686 zombie-eviction comment your diff had dropped, and pointed the tests at the real helper. Your root-cause analysis (restore in try instead of finally, lost because it lives on the per-turn event) was exactly right. Thanks!