Summary

• unify context_length override resolution for compression-related runtimes
• reuse custom_providers[].models[*].context_length across init, switch, fallback, and compression feasibility checks
• add regression coverage for override refresh/clear paths

Closes #8785.

Problem

Custom provider context length overrides were not applied consistently.

The main runtime could honor custom_providers[].models[*].context_length, but other paths could still re-probe provider metadata and fall back to 128000 when /models did not expose a context field. This caused spurious compression feasibility warnings and inconsistent runtime state.

Changes

• add a shared override resolver in AIAgent
• apply it during initial compressor setup
• apply it when switching runtimes
• apply it during compression feasibility checks for auxiliary compression models
• apply it when activating fallback runtimes and when restoring the primary runtime snapshot
• add regression tests for:
  • init uses custom provider override
  • switch refreshes override for new runtime
  • switch clears stale override when the new runtime has none
  • compression feasibility checks use the same override

Testing

• uv run --extra dev pytest tests/run_agent/test_switch_model_context.py tests/run_agent/test_primary_runtime_restore.py tests/agent/test_model_metadata.py -q -n 0