This PR bundles three related pieces of work for the CyberGym / GLM path in QitOS.

What changed:

• .gitignore ignores the locally-synced qitos/benchmark/cybergym/agent/ tree so private or local-only agent code does not leak into the public repo.
• qitos/benchmark/cybergym/runner.py switches CyberGym execution to runtime-budget-driven stopping, threads the correct workspace and task roots through the run, and keeps an internal step guardrail.
• qitos/recipes/benchmarks/cybergym.py exposes max_runtime_seconds at the recipe and CLI layer and makes max_steps optional so the runtime budget can drive execution.
• qitos/core/tool.py resolves annotations with get_type_hints and improves JSON schema type mapping for forward references, unions, optionals, and None.
• qitos/harness/_presets.py adds a dedicated glm family preset with native OpenAI-compatible tool-call preference and sensible protocol fallbacks.
• docs/reference/model-family-matrix.mdx and docs/zh/reference/model-family-matrix.mdx document the new GLM family row.
• tests/test_benchmark_cybergym_recipe.py verifies the recipe forwards runtime budget settings correctly.
• tests/test_tool_registry_and_toolset.py verifies annotated tool schemas resolve to valid JSON types.
• tests/test_harness_presets.py verifies GLM model names resolve to the glm family preset.

Validation:

• python -m unittest tests.test_benchmark_cybergym_recipe
• python -m unittest tests.test_harness_presets

Scope notes:

• This PR does not vendor cybergym_agent into QitOS.
• The CyberGym runner expects maintainers to sync the agent separately into qitos/benchmark/cybergym/agent/ when running the local benchmark recipe.