Bugfix: ignore legacy compression model for auto providers

[HiddenPuppy]

Summary

• ignore legacy compression.summary_model when compression.summary_provider is still auto
• stop passing that legacy model override into ContextCompressor, so auto-routed compression summaries use a provider-compatible default model
• update the example config and add regression coverage for both the config bridge and agent init path

Root Cause

The legacy compression section allowed a provider-specific model like google/gemini-3-flash-preview to remain configured while the provider stayed on auto. That value could be threaded into compression summary calls even when the resolved backend was not OpenRouter-compatible.

Validation

• pytest -n 0 tests/agent/test_auxiliary_client.py::TestTaskSpecificOverrides::test_task_without_override_uses_auto tests/agent/test_auxiliary_client.py::TestTaskSpecificOverrides::test_legacy_compression_model_ignored_when_provider_is_auto tests/agent/test_auxiliary_client.py::TestTaskSpecificOverrides::test_legacy_compression_model_preserved_for_pinned_provider tests/agent/test_auxiliary_client.py::TestTaskSpecificOverrides::test_explicit_compression_pin_still_wins_over_live_main_runtime tests/agent/test_auxiliary_client.py::TestTaskSpecificOverrides::test_compression_summary_base_url_from_config tests/run_agent/test_run_agent.py::TestInit::test_legacy_auto_compression_model_not_passed_to_context_compressor
• pytest -n 0 tests/agent/test_context_compressor.py
• pytest -n 0 tests/run_agent/test_413_compression.py

Closes #8923

[teknium1]

Addressed in #8992 which took this further — instead of patching around the legacy compression.summary_* keys, we removed them entirely along with the AUXILIARY_*/CONTEXT_* env var fallbacks. Config migration v17 auto-moves any existing values to auxiliary.compression.

Thanks @HiddenPuppy for identifying the bug and the clear root cause analysis in #8923 — that directly informed the fix.