[Tracking] March 2026 benchmark findings baseline

[Simonsbs]

Context

Latest benchmark snapshots (March 2, 2026) show mixed results:

• Apples-to-apples sum6 runtime: L0 501.81 Mops/s vs GCC 501.09 Mops/s (near parity).
• Build throughput on same kernel: L0 1250 ops/s vs GCC 60 ops/s.
• Real LLM cmd benchmark (qwen2.5:1.5b): verify 0%, semantic 0% across 6 tasks.
• Token compactness remains favorable: avg L0/C token ratio 0.6202.

Goal

Track this snapshot as the baseline and keep follow-up work linked to concrete deltas.

Tasks

• Add links to latest benchmark docs/wiki in this issue.
• Track improvement deltas for runtime throughput, build throughput, verify success, semantic success, and token ratio.
• Close only after successor tasks have landed and new benchmark reports are published.

References

• docs/PERFORMANCE_COMPARISON_APPLES_TO_APPLES.md
• docs/PERFORMANCE_COMPARISON.md
• docs/LLM_BENCHMARK_RESULTS.md