[codex] Add Snowl-native agent benchmark adapters

[ravenSanstete]

What changed

• Added Snowl-native evaluator primitives for answer matching, function-call matching, trace policy, command checks, workspace diffs, canary leakage, state transitions, checkpoint aggregation, grouped metrics, and normalized trace extraction.
• Added benchmark adapters/scorers for agent_bench_os, bfcl, ipi_coding_agent, and agentdojo.
• Registered the previously added safety/MCQ adapters and expanded benchmark docs so the README reflects the current built-in coverage.
• Added runtime support for per-sample dynamic tool schemas and a generic compose_terminal provider selected by runtime_container.provider_name.
• Native-ized toolemu scoring and examples so built-in code no longer bridges an external runtime package.

Validation

• python -m pytest -q -> 347 passed, 1 skipped
• python -m snowl.cli bench list
• Real remote API smoke with three models through Snowl:
  • deepseek-v3-ep
  • glm-5.1-w4a8
  • MiniMax-M2.7-w8a8
• End-to-end snowl bench run jsonl smoke: 3 trials, all success, remote_smoke_ok=1.000

Notes

• The PR intentionally leaves unrelated local changes unstaged, including docs.zip, webui/tsconfig.tsbuildinfo, and CLAUDE.md.