Document DeepSeek agent loop live eval

[SonAIengine]

Summary

• add the DeepSeek Flash 20-query agent-loop live eval artifacts
• update the public scale report with DeepSeek vs local qwen fallback comparison
• record DeepSeek-only hits, exploration behavior, latency/token trade-off, and delayed-discovery cases

Results

• DeepSeek Flash: 11/20 reach, zero-tool 0/20, duplicate calls 0, empty calls 8
• local qwen force-first fallback: 9/20 reach on the same subset
• DeepSeek-only hits: 178627, 68095, 1090242
• qwen-only hit: 1101278

Verification

• DeepSeek preflight: uv run python examples/ablation/run_agent_loop_benchmarks.py --llm-preset deepseek --preflight-only --preflight-timeout 15
• DeepSeek live eval: PYTHONUNBUFFERED=1 SYNAPTIC_SQLITE_FTS_AND_FIRST_THRESHOLD=20 SYNAPTIC_SQLITE_FTS_LEXICAL_RERANK_POOL=500 uv run python examples/ablation/run_agent_loop_benchmarks.py --llm-preset deepseek --subset 20 --msmarco-path tests/benchmark/data/msmarco_passage_full.json --corpus-limit 8841823 --sqlite-db-path tests/benchmark/data/msmarco_full.db --max-turns 5 --llm-timeout 180 --preflight-timeout 15 --out-jsonl examples/ablation/diagnostics/agent_loop_deepseek_v4_flash_20.jsonl --resume
• git diff --check