Benchmark provenance: hash a runner script at start, not at the end of the run

A per-run wrapper (bash run1.sh ..., started fresh for every cell by a driver loop) appended run1=$(sha256sum "$0") to its results row after the measured command finished. Mid-queue I replaced the script by temp file + mv, which is safe: the running bash keeps reading the old inode (check ls -l /proc/<pid>/fd/255 shows (deleted)) and the next cell picks up the new file. But the row of the cell that was in flight when the swap happened recorded the NEW script's sha, because the hash was taken at the end, from the path, not from the code that ran. Two rows ended up labelled with versions they never executed.

Fix: compute the hash once at the top, R1SHA=$(sha256sum "$0" | cut -c1-16) right after set -u, and write that variable into the row. For rows already written, map each cell to the version installed at its START time (keep an install log with UTC times) and record the mapping in a notes file next to the results instead of editing the results.

Related: once one script swap is approved, freeze the runner and the driver until a planned restart; every further swap re-opens the provenance question for whoever audits the rows.