fix(agent): clear stale turn state on model switch

[LeonSGP43]

Summary

• clear turn-scoped retry, prefill, and tool-guardrail state inside AIAgent.switch_model()
• prevent stale post-switch session state from leaking into the next API request
• add a regression test covering switch-time recovery-state reset

Testing

• ./venv/bin/pytest -q tests/run_agent/test_switch_model_fallback_prune.py
• ./venv/bin/ruff check run_agent.py tests/run_agent/test_switch_model_fallback_prune.py
• git diff --check

Attribution

• any repeated required-check red shared across recent company PRs is treated as preexisting_unrelated baseline noise; the local targeted proof for this diff is green.

Fixes #25325

[alt-glitch]

Related: #25343 (closed, not merged) — same fix pattern: clearing turn-scoped recovery state in AIAgent.switch_model(). Both address #25325.

Also related to #10702 (stale agent state on /resume), #25131 (persist model runtime state on switch).

[teknium1]

Automated hermes-sweeper review: this looks implemented on current main by the newer turn-prologue cleanup and retry-state refactor.

Evidence:

• agent/turn_context.py:131 resets the same turn-scoped state before each new turn: invalid/empty/prefill retry counters, post-tool fallback bookkeeping, mute/unicode state, tool guardrails, and _vision_supported.
• agent/conversation_loop.py:399 calls build_turn_context() before entering the API request loop, so the reset happens before the next request after /model switching.
• agent/conversation_loop.py:802 plus agent/turn_retry_state.py:33 also allocate a fresh TurnRetryState() per API-call attempt, covering the newer retry-guard state that used to live as scattered locals/flags.
• The PR's discussion linked this to [Bug]: Model switching corrupts session/request state with Hindsight enabled, causing retry loop and IndexError #25325/fix: reset model-switch recovery state and include follow-up cleanup #25343; [Bug]: Model switching corrupts session/request state with Hindsight enabled, causing retry loop and IndexError #25325's requested behavior was to clear or sanitize model-specific session/request state after switching, and current main now covers that path centrally.

Thanks for the focused regression test and for connecting it to the underlying stale-state report.