perf(bench): make a multi-hour acceptance run observable and restartable

The Phase-0 gate is 7 passes over the full row set, which is long enough
that capture_output=True was the wrong default: no rows existed anywhere
until a pass ended, so a slow run and a wedged one looked identical, and
killing either threw away everything measured.

Rows now stream to --out-dir/pass-NN.csv with a flush per line, so an
interrupted run keeps what it had. --resume reuses complete pass files
and reruns incomplete ones -- checked against the row count rather than
merely existing, because a partial file silently averaged in as a whole
pass would corrupt the verdict rather than fail it.

Progress is a tqdm bar over rows, not passes; a pass counter would sit
at 1/7 for twenty minutes and report nothing useful. The row total comes
from probing the binary with --reps=0 (0.8s) rather than reimplementing
the sweep in Python, which would drift from the C++ defaults. There is a
plain-stderr fallback when tqdm is absent -- refusing to start a
benchmark over a missing progress dependency is the wrong trade.

bench_runs/ is gitignored: the new default output path would otherwise
land in git status.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitignore

@@ -25,6 +25,9 @@ venv/
 .DS_Store
 Thumbs.db
 
+# Benchmark output (scripts/bench_repro_check.py --out-dir)
+bench_runs/
+
 # Claude Code local settings
 .claude/settings.local.json
 include/kpn/ort_cache/