What does this PR do?

Fixes a bug where _try_activate_fallback() calls get_model_context_length() without the config_context_length parameter. This causes the fallback activation to ignore the explicit model.context_length value in config.yaml and instead fall through to DEFAULT_FALLBACK_CONTEXT (128K).

Root cause

In run_agent.py, the switch_model() method (line 1988) correctly passes config_context_length=getattr(self, '_config_context_length', None) when calling get_model_context_length(). However, _try_activate_fallback() (line 6435) was missing this parameter.

The resolution chain in get_model_context_length() is:

• Step 0: If config_context_length is passed → return it immediately (respects config.yaml)
• Steps 1-8: Various probes (cache, endpoint, providers)
• Step 9: DEFAULT_FALLBACK_CONTEXT (128K) — the fallback for everything that fails

Without config_context_length passed, a fallback activation that fails all probes lands at step 9 → 128K, overwriting the correct 204800 value from config.

The crash cascade

1. Fallback model activates
2. get_model_context_length() called without config override → returns 128K
3. context_compressor.update_model(128K) overwrites the correct 204800
4. Status bar now shows 128K instead of 204.8K
5. Actual context grows past 128K → context exceeded error
6. Auto-compression fires on a corrupted session state → amnesia

Changes

run_agent.py — line ~6438

 fb_context_length = get_model_context_length(
     self.model, base_url=self.base_url,
     api_key=self.api_key, provider=self.provider,
+    config_context_length=getattr(self, "_config_context_length", None),
 )

This mirrors the existing correct behavior in switch_model() at line 1988.

Risk

Very low:

• One-line addition, identical to what switch_model() already does
• Only affects the fallback activation path (error recovery)
• Does not change behavior for successful model switches

Testing

• Manual: Switch to MiniMax-M2.7, trigger a fallback (e.g. API error), verify status bar still shows 204.8K not 128K
• Automated: Add a unit test that calls _try_activate_fallback() with a mocked _config_context_length and verifies get_model_context_length is called with that parameter

Fixes: context_length forced to 128K on fallback activation even when config.yaml has a higher value