Spike bench nits: corpus byte-count off-by-3, ruff format, report rows column

[valiantone]

Summary

Small accuracy/style nits found in local review of the spike bench code (#84). Non-blocking for merge; fix if the bench is ever re-run.

Findings

1. bench/native_spike/gen_corpus.py:171 — written += 38 but the malformed line is 41 bytes (40 + newline). Only skews the ~60% insertion threshold and final size by 3 bytes; bad_line_offset is recorded pre-write, so manifest correctness is unaffected.
2. bench/native_spike/run_bench.py fails ruff format --check (bench/ is not CI-enforced; one ruff format pass fixes).
3. bench/native_spike/run_bench.py:314 — the report's "rows" column takes the first arm's value blindly; parity is asserted separately, so cosmetic only.

Context

Local code review of #84 before merge; see bench/native_spike/README.md for the spike itself (#48).