fix: compression model should read context_length from custom_providers

[bestnite]

Problem

When auxiliary.compression.context_length is not explicitly set in config.yaml, the compression model's context length falls through to DEFAULT_CONTEXT_LENGTHS or the 128K fallback, skipping custom_providers[].models[].context_length entirely.

This means users who correctly configure their model's context length in custom_providers:

custom_providers:
- base_url: https://example.com/v1
  models:
    deepseek-v4-flash:
      context_length: 1000000

...still see the false warning:

⚠ Compression model (deepseek-v4-flash) context is 128,000 tokens

Fix

1. In init: Store normalized custom_providers on self._aux_compression_custom_providers
2. In _check_compression_model_feasibility(): After resolving aux_model and aux_base_url$, if no explicit context_lengthwas configured, match the compression model + base_url againstcustom_providers` entries — mirroring the main model's logic.

Fixes #8550

[alt-glitch]

Likely duplicate of #13813 — same fix: read context_length from custom_providers for compression model in run_agent.py. See also #13807 and #12977 for the underlying issue cluster.

[bestnite]

Agreed — #13813 covers this more broadly (incl. providers). Closing this one.