fix: Add tool-honesty instructions to prevent workers from fabricating results

[PureWeen]

Problem

Workers in multi-agent orchestration could fabricate results when their CLI tools failed to run, instead of honestly reporting the failure. The user expectation is: if tools don't run, workers should fail and report that — never fall back to self-evaluation or making up results.

Root Cause

Three prompt construction paths lacked tool-usage honesty instructions:

• Worker prompt (ExecuteWorkerAsync): No instruction about tool failures
• Synthesis prompt (BuildSynthesisPrompt): No guidance to verify tool usage in worker outputs
• Evaluator prompts (both BuildEvaluatorPrompt paths): No tool-verification scoring dimension

Fix

1. Worker prompt: Added CRITICAL: Tool Usage & Honesty Policy section requiring actual tool execution, honest failure reporting, and TOOL_FAILURE sentinel for unrecoverable errors
2. Synthesis prompt: Added instructions to flag fabricated results and preserve TOOL_FAILURE reports as-is
3. Evaluator prompts: Added Tool Verification as a 5th scoring dimension (multi-agent evaluator) and verification check (reflection cycle evaluator)
4. Synthesis-with-eval prompt: Added Tool Verification assessment dimension

Tests

Added 7 unit tests in WorkerToolHonestyTests.cs covering all modified prompt paths:

• Worker prompt contains tool-honesty instructions
• Synthesis prompt contains tool-verification guidance
• Synthesis prompt preserves TOOL_FAILURE reports
• Multi-agent evaluator includes Tool Verification dimension
• Synthesis-with-eval prompt includes Tool Verification
• ReflectionCycle evaluator includes tool verification
• ReflectionCycle evaluator preserves existing functionality

All 2544 passing tests continue to pass (4 pre-existing font-sizing failures unrelated).