fix(compression): reuse active model context for feasibility checks

[inside-ziwu]

Summary

• reuse the active model context when checking compression feasibility
• avoid feasibility decisions based on stale or mismatched context-length assumptions
• add regression coverage for the active-model path

Testing

• /root/.hermes/hermes-agent/venv/bin/python -m pytest tests/run_agent/test_compression_feasibility.py -q
• /root/.hermes/hermes-agent/venv/bin/python -m py_compile run_agent.py tests/run_agent/test_compression_feasibility.py

[alt-glitch]

Related to #8786 — both address context-length override propagation to compression paths. #8786 is broader (covers switch_model, fallback paths too).

[teknium1]

Thanks for identifying a real context-resolution mismatch. The premise still holds on current main: agent/context_compressor.py:1071-1075 initializes the active compressor with config_context_length, while the feasibility resolver independently calls get_model_context_length() in agent/conversation_compression.py:236-246 without reusing that active value.

Problems

• The changed run_agent.py body is no longer the live implementation. run_agent.py:1165-1168 now forwards to agent/conversation_compression.py, following extraction commit 5311d9959e19477aac8aa7deca46c1ee0b8e7000.
• Preserve the existing auxiliary-specific override when salvaging. agent/conversation_compression.py:240 passes _aux_compression_context_length_config, and agent/model_metadata.py:2051-2053 makes a positive explicit override highest priority. A same-runtime shortcut must not supersede it.

Suggested changes

• Port the narrow same-runtime fallback to agent/conversation_compression.py and retain the auxiliary override as the first choice.
• Update the regression coverage for both active-context reuse and auxiliary-override precedence.

This is an automated hermes-sweeper review.