Rebalance Pilot time and token budgets

[Chloride233]

Root cause

The one-task canary successfully created a Modal sandbox but spent about 107 seconds on first-run image, sandbox, Inspect tooling, model-proxy, and Codex setup. The frozen 90-second task limit expired before the first agent turn, yielding zero model usage.

Change

• reduce per-run token limit from 20K to 14K
• increase task limit from 90s to 180s
• increase sandbox timeout from 120s to 210s
• keep the same resource contract for canary and full Pilot
• update frozen budget evidence and documentation

Budget

• full 160-run conservative upper bound: ¥19.84384, below ¥20
• one-task canary conservative upper bound: ¥2.139524, below ¥2.25

Evidence

• python -m pytest -q: 293 passed
• python -m ruff check src tests benchmarks scripts: passed

The failed canary made zero DeepSeek calls and produced no attestation. This PR does not trigger another paid run.