The overnight tightening loop had a 99% error rate (11,058/11,210 tasks)
because resolve_runtime_provider() returned provider='local' which the
AIAgent doesn't recognize.
Fix: Bypass resolve_runtime_provider() entirely. The overnight loop
always runs against local Ollama inference — hardcode it.
Changes:
- Removed dependency on hermes_cli.runtime_provider
- Explicit Ollama provider (http://localhost:11434/v1)
- Model configurable via OVERNIGHT_MODEL env var (default: hermes4:14b)
- Base URL configurable via OVERNIGHT_BASE_URL env var
Before: 1% pass rate (139/11,210 over 1,121 cycles)
After: Should match Ollama availability (near 100% when running)
f8028cfb61