[aashish254]

What

Ten shipped error messages tell the user to run scripts/setup_benchmark.py <benchmark>, but the
file is committed mode 100644, so typing that line ends in a kernel rejection instead of setup:

$ scripts/setup_benchmark.py jb
env: scripts/setup_benchmark.py: Permission denied
$ echo $?
126

This PR sets the executable bit on that one file -- the diff is old mode 100644 / new mode 100755, no content line changes (871d1cb9… before and after, sha256sum 80291a59cf6e…
identical in both trees).

The sites, counted over all tracked .py and .sh files:

Form	Count	Where
run scripts/setup_benchmark.py … (bare path)	10	benchmarks/jobbench.py:970,1081, benchmarks/lab.py:428,473, benchmarks/workbuddy.py:945, core/proposer.py:402, substrate/codex.py:219, substrate/opencode.py:313, third_party/jb_runner/run_benchmark_codex_cli_adapt.sh:493,494
python scripts/setup_benchmark.py …	4	jobbench.py:981, lab.py:936,958, workbuddy.py:959

One of the ten is the line a first-time deployer hits first (harness2/substrate/opencode.py:311-314):

OpenCode checkout missing at <root>/dependencies/opencode (no packages/opencode/src/index.ts):
run scripts/setup_benchmark.py <benchmark> to fetch the pinned checkout at 23fb5e0516c9 (v1.14.18).

Why the mode bit rather than rewording the messages

• Two of the ten live in third_party/jb_runner/run_benchmark_codex_cli_adapt.sh, which is
  covered by a dependencies/ASSETS.json checksum -- editing it means editing ASSETS.json, and
  8 open PRs are already fighting over that file. The mode bit makes those two readable without
  touching them.
• The other eight are spread over six modules that 20 distinct open PRs are editing at
  neighbouring lines (jobbench.py alone is in fix(jobbench): parse run identity from its path suffix #18/fix(jobbench): keep judge worker exceptions ungraded #28/fix: stop passing API keys and host tokens to agent subprocesses #37/Add Claude Code (cc) and Pi (pi) substrates for JobBench #66/core: include task inputs in the reuse identity of a rollout #76/feat: support saved ChatGPT login for native Codex runs #81/Let Claude Code authenticate without Vertex AI; selectable claude_cli improver #84/Add a claude_cli JobBench judge backend #85/fix(jobbench): a judge that produced no total is ungraded, not a zero #89/fix(jobbench): clean up render directory after rollout #90). A
  wording sweep there is a merge-conflict generator; a one-file mode change is not.
• The repo's own convention supports the distinction: docs/EXPERIMENTS.md:59-61 and the
  scripts/run_*.sh headers always invoke those wrappers as bash scripts/run_lab.sh, so their
  100644 is correct and left alone. Only setup_benchmark.py is told to users as a path to
  execute, and only it gets the bit.
• The console-script form (harness2-setup jb, which README and docs/SETUP.md document) is of
  course already fine; this fixes the script-path form the product itself prints.

Verification

Python 3.12.13, measured in a pristine 1cfe254 worktree (/tmp/h2base) and a worktree carrying
only this mode change (/tmp/h2exec). The venv PATH stands for README's documented steps 1-3
(uv venv --python 3.12 .venv && source .venv/bin/activate && uv pip install -e '.[render]').

#	Environment	Invocation	rc	Output
A	base 100644, venv PATH	scripts/setup_benchmark.py --help	126	env: scripts/setup_benchmark.py: Permission denied
B	this change 100755, venv PATH	scripts/setup_benchmark.py --help	0	argparse usage, Install a benchmark's pinned tools, runner overlays, dependencies, and data.
C	this change, venv PATH	./scripts/setup_benchmark.py --help	0	same usage
D	base, venv PATH	python scripts/setup_benchmark.py --help	0	same usage -- the 4 python-prefixed messages already work, which is why only the bare sites are broken

Rows A vs B differ in nothing but the mode bit, so the fix is proven at the exact seam the message
describes. Shebang resolution was checked rather than assumed: inside the activated venv
/usr/bin/env python3 is Python 3.12.13, which is inside requires-python = ">=3.12,<3.14".

Everything else is unchanged, as it must be for a content-free diff:

Check	base	this change
python -m unittest discover -s tests	Ran 5 tests / OK	Ran 5 tests / OK
ruff check .	rc=0	rc=0
git diff --stat 1cfe254..HEAD	--	1 file changed, 0 insertions(+), 0 deletions(-)

What this does not fix, stated plainly: with the bit set but no venv activated, the script
still fails -- env PATH=/usr/bin:/bin ./scripts/setup_benchmark.py --help gives rc=1 and
ModuleNotFoundError: No module named 'harness2', because the system interpreter here is 3.9.6
and the package is not installed. python scripts/setup_benchmark.py fails identically at base,
so this PR neither adds nor removes that constraint; the venv is the only supported environment
either way.

No test is added on purpose: a os.access(..., X_OK) assertion would pass or fail depending on
the checkout's platform and core.fileMode, not on the code under review -- a Windows checkout or
a tarball without mode bits would report red on a correct tree. The A/B measurement above is the
witness.

Not a duplicate

• No issue describes it: all 23 issues (open + closed, Proposal: opt-in human feedback for selection and lessons (off by default) #2..How to use with Antigravity #100) were searched for
  permission denied, setup_benchmark, executable. The hits are incidental uses of the word
  "execute" in unrelated reports.
• gh pr list --state open --limit 200 --json files over all 75 open PRs: 0 touch
  scripts/setup_benchmark.py. The only PRs under scripts/ are fix(workflow): propagate benchmark report failures #9 and fix(workflow): preserve seed across refinement and grading #17, which edit
  run_jobbench.sh / run_lab.sh / run_workbuddy.sh (untouched here), and my own style(scripts): make render_readme_typing.py satisfy ruff format #103, which
  reformats scripts/render_readme_typing.py -- a different file, so no overlap.
• git merge-tree --write-tree onto main, in both orders: clean. Re-run against the two
  neighbouring scripts/ PRs at their current heads (fix(workflow): propagate benchmark report failures #9 e72bb12, fix(workflow): preserve seed across refinement and grading #17 d1bf6f0): clean both
  ways, and git diff-tree --summary shows neither changes a file mode. (My first pass reported
  these as conflicts only because the refs had not been fetched yet -- corrected by fetching and
  re-measuring, not by reasoning about it.)
• Disclosure: ruff format --check . reports 1 file would be reformatted, 49 files already formatted in both trees, for scripts/render_readme_typing.py. That is pre-existing on main
  and is what PR style(scripts): make render_readme_typing.py satisfy ruff format #103 fixes; it is not introduced or affected here.