Summary

• parse OpenRouter/Nous N in the output context errors as output-cap errors
• compute the available output budget from context length minus text/tool input
• add regression coverage using the Nous error format from [Bug]: parse_available_output_tokens_from_error() misses OpenRouter/Nous "in the output" format — causes infinite auto-reset loop #38652

Fixes #38652

Tests

• python -m pytest tests/test_ctx_halving_fix.py::TestParseAvailableOutputTokens::test_openrouter_nous_output_tokens_format -q -o addopts= --tb=short
  • failed before the fix with None == 236907
  • passes after the fix
• python -m pytest tests/test_ctx_halving_fix.py -q -o addopts= --tb=short
• python -m py_compile agent/model_metadata.py tests/test_ctx_halving_fix.py
• ruff check agent/model_metadata.py tests/test_ctx_halving_fix.py
• git diff --check