Preserve paired array call boundaries

[SonAIengine]

Summary

• Tighten BFCL model-loop instructions so array-typed parameters do not cause distinct item/time pairings to be merged into one tool call.
• Add array argument description guidance that tells the model to keep only one pairing per call and emit another call for the next pairing.
• Recover BFCL parallel_9 for both row-source and retrieved-source Qwen smokes.
• Update research docs with before/after artifacts and mark multiple_7 as the remaining extra-related-tool-call bottleneck.

Validation

• poetry run ruff check .
• poetry run ruff format --check .
• poetry run pytest tests/test_bfcl_tool_selection_llm_loop.py -q -> 27 passed
• poetry run pytest tests/ -q -> 969 passed, 3 skipped, 1 warning
• Qwen before row/retrieved: /tmp/gtc-bfcl-repeated-before-row-qwen.json and /tmp/gtc-bfcl-repeated-before-retrieved-qwen.json -> parallel_9 call_count_mismatch
• Qwen after row: /tmp/gtc-bfcl-paired-array-after-parallel9-row-qwen.json -> exact 1.00
• Qwen after retrieved: /tmp/gtc-bfcl-paired-array-after-parallel9-retrieved-qwen.json -> exact 1.00
• Qwen recovered 3-case smoke: /tmp/gtc-bfcl-paired-array-recovered3-qwen.json -> 3 / 3 pass, exact 1.00

Remaining Bottleneck

• multiple_7 still fails from an extra related analysis call even with row-source tools; that is the next model/schema-expression issue to investigate.

[SonAIengine]

Closing this internal stacked PR as superseded by the current main branch. The branch is preserved for history.