ASI feedback + GEPA integration + unified run.py

[ashvin-verma]

Summary

• ASI (Actionable Side Information): 5-tier structured feedback for LLM-guided LLVM heuristic evolution
  • Tier 1: Score decomposition + per-benchmark signal classification (always on)
  • Tier 2: Compiler statistics delta via -stats (default on)
  • Tier 3: Runtime variance analysis (always on)
  • Tier 4: Hardware perf counters via perf stat (opt-in)
  • Tier 5: Optimization decision changes via -pass-remarks-output YAML diff (opt-in, ~20% overhead)
• GEPA integration: Alternative to OpenEvolve using Google DeepMind's optimize_anything(). ASI feeds directly into GEPA's native side-info channel as (score, {"Feedback": asi_text})
• Unified run.py: Single CLI dispatches to GEPA (default) or OpenEvolve via --framework. Replaces 4 separate files (manual_run.py, gepa_run.py, gepa_adapter.py, providers.py) with 3 clean modules (run.py, adapters.py, evaluator.py)

Architecture

run.py --framework {gepa, openevolve} --task {llvm_inlining, ...}
  │
  ├─ GEPAAdapter  (Pareto frontier, native side-info)
  └─ OpenEvolveAdapter  (MAP-Elites population)
       │
       ├─ ManualLM / ManualLLM  (file-based prompt/response)
       └─ evaluator.py → tasks/*/evaluate.py → llvm_bench.py
            (patch LLVM → build → benchmark → Optuna → ASI)

Test plan

• run.py --framework gepa --task llvm_inlining --max-evals 2 --auto — GEPA smoke test with real LLVM builds (2 evals, seed + auto-response)
• All Python files compile clean (py_compile)
• No dangling imports to deleted files
• README updated with unified CLI usage, architecture diagram, comparison table

🤖 Generated with Claude Code