Add model-in-loop long-horizon benchmark

[SonAIengine]

Summary

• add a real model-in-the-loop gate above the sealed 1,000-tool Arazzo 3/10/30-step benchmark
• separate raw LLM target accuracy from selector-guard and end-to-end goal completion
• add explicit DeepSeek provider options, credential-safe endpoint redaction, Make target, tests, and public result artifact

Actual model result

DeepSeek V4 Flash, 3 repeats across 3/10/30-step scenarios (9 executions):

• retrieval target hit@8: 1.00
• raw LLM target exact: 1.00
• final selected target exact: 1.00
• exact plan/execution order: 1.00 / 1.00
• binding accuracy: 1.00
• goal completion: 1.00

The model selects only the final target from the retrieved catalog. Arazzo supplies prerequisite order and bindings; the benchmark does not claim independent 30-call enumeration by the model.

Verification

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/ -q (1244 passed, 47 skipped)
• DeepSeek V4 Flash 1,000-tool 3x model gate: pass
• saved artifact secret scan: pass