bench: add generic runner and M1 Max Qwen3.6 results

[kira-ariaki]

Summary

• add the generic benchmarks/scripts/run_bench.sh entrypoint already referenced by the benchmark docs and manual workflow
• parse basert-bench output fail-closed into normalized CSV with engine-free tests
• run benchmark-tool checks in CI
• publish BaseRT 0.1.7 Qwen3.6-35B-A3B default-q4 results from an M1 Max 64 GB
• preserve model/source SHA-256 provenance, converter details, and complete raw second-run output
• clarify independent prefill/decode microbenchmark semantics

Results contributed

Five measured repetitions after one warm-up iteration:

• pp128: 552.07 tok/s
• pp256: 692.73 tok/s
• pp512: 851.39 tok/s
• pp2048: 1025.34 tok/s
• pp6000: 878.60 tok/s
• pp43000: 574.93 tok/s
• tg128: 95.99–97.15 tok/s across the six independent invocations

These are single-stream engine microbenchmarks, not server throughput or decode-after-prefill measurements.

Provenance

• source revision: Qwen/Qwen3.6-35B-A3B@995ad96eacd98c81ed38be0c5b274b04031597b0
• conversion: basert pull Qwen/Qwen3.6-35B-A3B, base-convert v0.1.6, profile default-q4
• .base SHA-256: 74c4a9256e64971e919c41b78b03ff8e0ec74b4330058113f5785097706ed1d0
• BaseRT 0.1.7 basert-bench SHA-256: f81d0ddd3c3eae1ff43f734aa0c76786b8b2e9b9fe3e1c5ec28f8d581a972721

Verification

• python3 -m unittest discover -s benchmarks/tests -v — 5/5 passed
• ruff check benchmarks/scripts/parse_bench.py benchmarks/tests
• python3 -m py_compile benchmarks/scripts/parse_bench.py benchmarks/tests/*.py
• bash -n benchmarks/scripts/run_bench.sh
• reran the complete six-point sweep end-to-end through the new runner; all throughput values reproduced within 1.2%, and the committed CSV/raw output record that second run