Runtime compression status and artifact extraction contract

[1aal]

Summary

• keep PR Runtime compression status and artifact extraction contract #21 wrapper-only: no changes to native Hermes Agent core files under agent/* or run_agent.py
• codify Runtime worker LLM config resolution so Cloud run config with endpoint credentials is normalized to Hermes custom with the same model/base_url/api_key; this covers both explicit openai-compatible and display-provider names such as Qwen when base_url + api_key are present
• leave model limit entries in the Runtime wrapper contract: max_tokens is passed into AIAgent, and context_length is applied to the agent/compressor config so Hermes native compression can use the configured context window
• project existing Runtime status messages about context compression degraded/failed into structured JSONL events in Runtime worker so Cloud can mark warning/partial without reading pod logs
• preserve report artifact metadata/download behavior: no local paths and no full artifact content in turn events

Notes

• This does not change native Hermes Agent compression behavior. It surfaces existing compression warnings from the Runtime wrapper layer and feeds runtime config into existing Agent/compressor paths.
• P0 auxiliary LLM policy: if Cloud provides a complete run llm_config, Runtime should make that provider/model/base_url/api_key available to the Agent and compression auxiliary path; if unavailable, Runtime emits structured compression degraded/failed events for Cloud to display as warning/partial.
• Provider display names must not decide compression routing when Cloud provides an explicit endpoint. For example, provider=Qwen plus baseURL/apiKey is routed as custom + base_url + api_key so auxiliary compression reuses the same Cloud run endpoint instead of Hermes' built-in Qwen/Alibaba provider path.
• Supported limit fields are canonical max_tokens and context_length; camelCase aliases are accepted from the worker request for compatibility, and context_window aliases are accepted only as migration compatibility into canonical context_length.
• This PR intentionally does not provide a Runtime /extract API or bounded report-content extraction. Report summarization should rely on Hermes native context compression, not a separate Cloud/Runtime summary feature.
• Cloud still needs to project the new structured compression events.

Verification

• uv run --extra dev pytest tests/runtime_manager/test_registry.py::test_runtime_worker_treats_named_cloud_provider_with_endpoint_as_custom tests/runtime_manager/test_registry.py::test_runtime_worker_resolves_model_limit_entries_from_llm_config tests/runtime_manager/test_registry.py::test_runtime_worker_applies_context_length_to_agent_compressor tests/runtime_manager/test_registry.py::test_runtime_worker_resolves_cloud_llm_config_for_agent_and_auxiliary tests/runtime_manager/test_registry.py::test_runtime_worker_normalizes_cloud_provider_aliases_to_hermes_names tests/runtime_manager/test_registry.py::test_runtime_worker_projects_compression_status_as_structured_event tests/runtime_manager/test_report_artifacts.py -q
• uv run --extra dev ruff check runtime_manager/worker_main.py tests/runtime_manager/test_registry.py
• python3 -m py_compile runtime_manager/worker_main.py
• git diff --check
• git diff --check origin/apecloud-base...HEAD

Related to task NousResearch#111 / issue apecloud/apecloud#18712.