fix: honor custom provider context for auxiliary compression

[cherished-liuxu]

Summary

• preserve the custom provider context when deriving the auxiliary compression model
• avoid falling back to the main provider when large is configured under a custom provider
• add a regression test covering the custom-provider auxiliary compression path

Test Plan

• pytest -q tests/test_aux_compression_custom_provider_context.py

[alt-glitch]

Duplicate of #13813 (and competing with #8721, #21953, #8786). Multiple open PRs address the same bug: auxiliary compression feasibility check not threading custom_providers through to get_model_context_length(). See also tracking issue #12977.

[teknium1]

This appears to be implemented on main already. This is an automated hermes-sweeper review.

Evidence:

• agent/conversation_compression.py:136 now calls get_model_context_length() for auxiliary compression with custom_providers=agent._custom_providers.
• agent/model_metadata.py:1529 / agent/model_metadata.py:1565 show get_model_context_length() accepts custom_providers and checks custom provider per-model context overrides before fallback probing.
• The fix was introduced by 7becb19ea00c13bdff6f78b71aa3ddfb0bdb5378 (fix(auxiliary): forward custom_providers to compression model context-length detection) and is contained in release tag v2026.5.16.
• The maintainer discussion on this PR identified it as a duplicate of the same missing custom_providers threading bug.