Document DeepSeek agent loop 50 query eval

[SonAIengine]

Summary

• add the DeepSeek Flash 50-query agent-loop eval artifacts
• update the public scale report with the 50-query summary row
• record delayed-discovery hits and high-call miss bottlenecks for next tuning

Results

• DeepSeek Flash 50-query check: 23/50 reach
• zero-tool answers: 0/50
• duplicate tool calls: 0
• empty tool calls: 11
• delayed discovery hits: 178627, 87892, 1090242, 45924, 323998, 333486
• high-call misses: 54544, 293992, 208145, 14151, 91711, 237373

Verification

• DeepSeek 50-query live eval with --llm-preset deepseek, --subset 50, --resume
• rg -n "DEEPSEEK_API_KEY=|sk-" examples/ablation/diagnostics/agent_loop_20260702_201106.md examples/ablation/diagnostics/agent_loop_deepseek_v4_flash_50.jsonl examples/ablation/diagnostics/public_scale_20260702.md || true
• git diff --check