feat: extend benchmarks with agentic variation mode support and comparison

[danielbentes]

Summary

Extends the benchmark infrastructure to support the hybrid agentic evolution feature (v1.1.0) and adds a comparison benchmark script for evaluating all 3 variation modes.

• Config wiring: SelfImprovementConfig and EvolutionBenchmarkConfig now accept AgenticVariationConfig, threaded to EvolutionScheduler
• CLI flags: --variation-mode, --max-inner-iterations, --agent-model added to run_self_improvement_benchmark.py
• Comparison script: New run_agentic_comparison.py runs all 3 modes (single_turn, agentic, adaptive) on the same dataset
• Comparison report: New AgenticComparisonReport generates side-by-side analysis with winner analysis and recommendations
• Agentic metrics tracking: GenerationSnapshot.agentic_stats captures inner iterations and budget per generation
• Report updates: Self-improvement reports include variation mode in metadata and configuration

Usage

# Quick comparison (3 generations, 20 samples)
python -m siare.benchmarks.scripts.run_agentic_comparison \
    --provider ollama --model llama3.1:8b \
    --reasoning-model deepseek-r1:7b \
    --dataset-tier 1 --quick

# Full comparison
python -m siare.benchmarks.scripts.run_agentic_comparison \
    --provider openai --model gpt-4o-mini \
    --reasoning-model gpt-4o \
    --dataset-tier 1 --generations 10 --samples 50

# Single mode via existing benchmark
python -m siare.benchmarks.scripts.run_self_improvement_benchmark \
    --provider ollama --model llama3.1:8b \
    --variation-mode adaptive --quick

New files

File	Purpose
siare/benchmarks/scripts/run_agentic_comparison.py	CLI for running all 3 modes
siare/benchmarks/reports/agentic_comparison_report.py	Side-by-side comparison report
tests/unit/test_benchmark_agentic_config.py	14 tests for config wiring and reports

Test plan

• 389 tests pass (375 existing + 14 new), 0 regressions
• ruff check siare/ — all checks passed
• Backward compatible — all defaults are None (existing behavior)
• Run comparison benchmark with real LLM provider

🤖 Generated with Claude Code

[danielbentes]

Self-Review: PR #8

Issues Found and Fixed (commit bbec2f6)

P1 Fixes (4 — were blocking)

#	Issue	Fix
1	AgenticComparisonReport(mode_results) missing 2nd arg + wrong type	Pass (mode_summaries, mode_configs_dict)
2	report.save_markdown() method doesn't exist	Changed to report.save(str(output_dir))
3	metric_data.get("final") — key is "evolved"	Fixed key name
4	(Combined with #1) — dataclass objects instead of dicts	Used mode_summaries

P2 Fixes (4)

#	Issue	Fix
1	+-5.0% for negative improvements	Use {pct:+.1f}% format
2	finish_reason hardcoded to "stop"	Extract done_reason from Ollama response
3	Default model llama3.2:7b doesn't exist	Changed to llama3.2:3b
4	Docs config says gpt-4o, actual default is gpt-5	Fixed YAML example

P3 Fixes (1)

• Catch ValueError for JSON parse errors in OllamaProvider

Remaining P2 (deferred)

• agentic_config: Any type safety — requires dataclass/Pydantic compatibility work
• check_ollama_model() substring match — minor, existing behavior

Quality Gates

• 389 tests pass
• Lint clean
• All P1s fixed

🤖 Generated with Claude Code