fix: boost output tokens near max turns for completion

[drewstone]

Summary

• When the agent is within 4 turns of max, bump maxOutputTokens from 600 to 1200
• Data-heavy completion responses (multi-record extractions) truncate at 600 tokens, causing malformed JSON that wastes remaining turns on retry loops
• Most turns use 200-350 output tokens regardless of limit — the higher cap only costs more when the LLM actually needs it
• Fixes webbench-108 (Apple store locator) where agent successfully extracted all data by turn 16 but then crashed turns 17-20 on truncated completion JSON

Test plan

• pnpm build passes
• Run webbench-108 (Apple store locator) to verify completion succeeds
• Verify token cost doesn't increase significantly on typical cases (most output is well under 600)