DI-4855 Add Deep Agents Code benchmark runner

[wiggzz]

Why

We need a directly comparable ADE-Bench baseline for LangChain's Deep Agents Code (dcode) while evaluating Deep Agents for Wizard.

What

Adds dcode as a first-class installed ADE-Bench agent, pinned to deepagents-code==0.1.45. It runs headlessly with explicit model/reasoning configuration, reuses the benchmark AGENTS.md, preserves transcripts, records runtime, and parses the CLI usage table into approximate token/request metrics.

dcode's headless output rounds compact token values and does not expose cache-token or cost breakdowns, so token totals are explicitly approximate and cost remains unavailable.

Validation

• Adapter tests, including public task execution and usage parsing: 3 passed.
• Ruff and Black checks: passed.
• Full suite: 85 passed; one unrelated test hard-codes allowed checkout directory names and rejects this shared checkout name.
• Completed the GPT-5.6 Sol, low-reasoning, 75-task × 3-attempt run: 167/225 resolved (74.2%), two agent timeouts, ~78.46M reported input tokens, ~570.9K output tokens.
• The paired Codex run resolved 165/225 (73.3%); dcode was better on 12 task-level three-attempt scores, Codex on 10, and 53 tied. dcode's evaluation took 1h40m25s versus Codex's 1h09m25s.

Refs

• DI-4855
• ai-codegen-api integration PR
• Completed dcode GPT-5.6 Sol 3× run
• Completed Codex GPT-5.6 Sol 3× run
• Earlier dcode GPT-5.4 run
• Deep Agents CLI documentation

Drafted by Codex (GPT-5) under direction of @wiggzz.