mbo/hash/measurements: one-shot authoritative runner + parallel SMHasher batteries

[helly25]

Dev-only tooling under the .bazelignore'd mbo/hash/measurements module - no shipped code changes. Makes the authoritative measurement run a single command and cuts the SMHasher3 wall-clock.

run_measurements.py (new)

One command, run from a clean main checkout, that:

1. runs the full performance sweep (solo - it is sub-ns and SMHasher3 would contend for CPU),
2. renders the ns-vs-length chart,
3. runs the SMHasher3 battery (parallel),
4. promotes the canonical JSON + SVG to stable committed names and prints exactly what to commit.

mbo/hash/measurements/run_measurements.py --jobs 4          # everything
mbo/hash/measurements/run_measurements.py --algos mumbo,jumbo,dumbo --jobs 1
mbo/hash/measurements/run_measurements.py --skip-smhasher   # perf + chart only

hash_benchmark_report.py smhasher gains --jobs

The per-algorithm batteries are independent, and their pass/fail verdicts are load-independent (only SMHasher3's own Speed sub-test would be affected by contention, and that number is unused - performance comes from our own benchmark). So --jobs N runs several concurrently, trading cores for wall-clock (--algos all drops from ~2 h to well under 1 h).

--smhasher3 is now a command prefix (shell-split), so it works both as a native binary path and as a docker run ... ./build/SMHasher3 wrapper. That's needed on macOS, where the container-built Linux binary can't run natively; the runner constructs the wrapper automatically (image matches build_smhasher3.sh).

Notes

• The committed dataset is only tagged authoritative from a clean main tree; the runner warns otherwise.
• data/ stays .gitignored; the runner prints the git add -f commands for the raw + SMHasher artifacts, and the plain git add for the canonical JSON + chart.
• Measured artifact sizes are repo-friendly: raw ~316 KB gzip, canonical ~100 KB, SVG ~16 KB, each SMHasher log ~20 KB gzip.