fix: compaction fallback to lowest model context limit

[shaj13]

Problem

When the provider is not found in the models store, compaction fails to determine the correct context limit for the model.

providers:
  xxx:
    api_type: openai_chatcompletions
    base_url: xxxxx

models:
  claude-opus-4-5:
    provider: xxx
    model: claude-opus-4-5
..... 

Fix

Added a fallback that searches all providers for the model and uses the lowest context limit found.
This ensures compaction triggers conservatively regardless of which provider is actually being used at runtime.