bench: add reps (median ± spread) to compare_bench.py

[cjfields]

Summary

Makes within-node benchmark noise visible. A label's PATH can now be a
comma-separated list of rep CSVs/dirs — repeated runs of the same config —
reduced to the median per metric, with the relative half-range shown as
median ±N%. Δ% compares medians.

Motivated by this investigation: a scary "+19% wall" turned out to be
concurrent-job contention on a shared node, not code. Reps surface that kind of
run-to-run variance directly instead of letting a single noisy run mislead.

Changes

• parse_run: LABEL=PATH[,PATH...] — each path a CSV file or a directory.
• load_reps: median + spread (relative half-range) per (key, metric) across reps.
• print_table: median ±N% cells when reps>1; banner shows label (n=N).
• --json: emits {nreps, median, spread} per run.

Single-path usage is unchanged (no spread shown). Pure stdlib; benchmark tooling only.

Example

compare_bench.py --metric maxrss_kb \
    --baseline main=r1_main,r2_main,r3_main \
    --compare branch=r1_br,r2_br,r3_br

### maxrss_kb  (↓ better)
                     main              branch      Δ%
dada         13.25 GB ±2%         8.80 GB ±2%  -33.6%

🤖 Generated with Claude Code