Support pre-defined multi-turn conversation history in BenchmarkCase

[shanevcantwell]

Context

BenchmarkCase currently supports single-turn only (system + user → one response). For ReAct loop evaluation (ADR-012 Phase 1), we need to support pre-defined multi-turn conversation histories so the model responds to the final message with full accumulated context visible.

This enables:

• Archive replay from LAS production traces (serialize first N iterations as context, test model's iteration N+1)
• Context degradation testing without a full loop executor
• Foundation for dynamic ReAct evaluation in later phases

Changes

• Add optional messages: list[dict] field to BenchmarkCase
• Update to_messages() to return messages if present, else fall back to system + user
• No changes needed to runners, MCP tools, or adapters — they already accept list[dict] messages

Acceptance

• Tests for to_messages() with and without messages field
• Backward compatibility: all existing 331 tests pass
• JSON battery file with messages array loads and runs

Part of ADR-012: ReAct Loop Evaluation (Phase 1 of 6)