Fix context compaction failures on temperature-restricted models

[MadKangYu]

Summary

• stop forcing a fixed temperature in context summary LLM calls
• add a summary failure cooldown to avoid repeated compaction spam on persistent provider errors
• downgrade repeated no-summary log noise to debug
• add tests for no-temperature kwargs and cooldown behavior

Validation

• /Users/yu/.hermes/hermes-agent/.venv/bin/python -m pytest -q /Users/yu/.hermes/hermes-agent/tests/agent/test_context_compressor.py
• 36 passed

[MadKangYu]

Follow-up context for reviewers:\n\nRepro seen in production logs:\n- context compaction repeatedly triggers summary generation\n- summary call fails with: \n- repeated failure leads to compaction retry loop + warning spam\n\nWhat this PR changes:\n1. Stops forcing a fixed temperature in context summary calls (lets provider/model defaults apply)\n2. Adds a short failure cooldown for summary generation to prevent immediate repeated failures\n3. Downgrades repeated no-summary noise to debug\n\nWhy this is safe:\n- If summary generation fails, compaction still proceeds without summary (existing fail-open behavior), but now avoids spam/retry storms.\n- Includes unit tests for no-temperature kwargs and cooldown behavior.\n\nLocal validation run:\nbringing up nodes...
bringing up nodes...

.................................... [100%]
=============================== warnings summary ===============================
tests/agent/test_context_compressor.py: 36 warnings
/Users/yu/.hermes/hermes-agent/tests/conftest.py:89: DeprecationWarning: 'asyncio.get_event_loop_policy' is deprecated and slated for removal in Python 3.16
loop = asyncio.get_event_loop_policy().get_event_loop()

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
36 passed, 36 warnings in 3.90s\n=> 36 passed

[MadKangYu]

Correction (formatted cleanly):

Repro seen in production logs:

• context compaction repeatedly triggers summary generation
• summary call fails with: invalid temperature: only 1 is allowed for this model
• repeated failure leads to compaction retry loop + warning spam

What this PR changes:

1. Stops forcing a fixed temperature in context summary calls (lets provider/model defaults apply)
2. Adds a short failure cooldown for summary generation to prevent immediate repeated failures
3. Downgrades repeated no-summary noise to debug

Why this is safe:

• If summary generation fails, compaction still proceeds without summary (existing fail-open behavior), but now avoids spam/retry storms.
• Includes unit tests for no-temperature kwargs and cooldown behavior.

Local validation run:
/Users/yu/.hermes/hermes-agent/.venv/bin/python -m pytest -q /Users/yu/.hermes/hermes-agent/tests/agent/test_context_compressor.py
=> 36 passed