Setting a local model as primary rewrites cascade rows to the old primary's provider (drops ollama base_url)

[juniperbevensee]

Observed on Matilde (h_matilde) 2026-07-05. Cascade before the change (correct, per fleet-wide GLM rollout): anthropic/fable -> anthropic/opus -> ollama/glm4:9b (base_url) -> ollama/qwen3:30b (base_url).

After switching the primary to glm4:9b in the HSM models UI, config.yaml came out as:

model:
  provider: anthropic     # <- old primary's provider kept
  default: glm4:9b
fallback_providers:
  - {provider: anthropic, model: glm4:9b}    # <- base_url dropped
  - {provider: anthropic, model: qwen3:30b}  # <- base_url dropped
  - {provider: anthropic, model: claude-fable-5}

Every request then went to the Anthropic API with model=glm4:9b -> 404 -> 'provider failed after retries' in chat. Fallback rows could not rescue it because they were rewritten to anthropic too.

Same family as #126 (models-PUT col-0 corruption). Suspect: the UI/route derives provider for all rows from the harness's top-level provider instead of preserving each row's provider+base_url when reordering the primary.

Workaround applied: PUT /api/harnesses/h_matilde/models with explicit fallback_providers (ollama+base_url rows) + quick restart — works.

🤖 Generated with Claude Code