JVNAUTOSCI-1428 gate local Ollama fallback probes

[witbrock]

Summary

• reduce the default Ollama reachability probe timeout to 200ms for local hosts while keeping 1200ms for non-local hosts
• preserve cached failure duration separately so cooldown hits report near-zero reuse time instead of replaying the original timeout
• add targeted regressions for cooldown-cache telemetry and local-vs-remote timeout selection, using a lightweight orchestrator fixture to avoid unrelated startup cost in the test file

Verification

• pdm run pyright src/backend/integrations/internal_mcp/orchestrator.py tests/backend/test_orchestrator_model_fallback_execution.py
• pdm run pytest tests/backend/test_orchestrator_model_fallback_execution.py -q
• pdm run pytest tests/backend/test_orchestrator_workflow_selector_routing.py::test_workflow_selector_uses_provider_aware_classifier_fallback -q
• pdm run pytest tests/backend/test_tool_progress_liveness.py::test_progress_clears_stale_error_on_subsequent_success -q
• attempted pdm run pytest tests -m lane_backend_mcp -q but it exceeded the agent timeout in this environment