Selftest wipes PONYTAIL_PLUGIN_DIR before the run can use it

[KimHyeongRae0]

I was re-running benchmarks/agentic to check the numbers in results/2026-06-18-agentic.md — different machine, plugin not installed there, so I set PONYTAIL_PLUGIN_DIR the way benchmarks/agentic/README.md suggests. Baseline cells ran fine, then the first ponytail cell died:

ponytail plugin dir not found under ~/.claude/plugins/cache/ponytail/ponytail; install the plugin or set PONYTAIL_PLUGIN_DIR

Which was confusing, because I had.

_selftest_plugin_dir() sets the var to a sentinel to prove the override is honored, then unconditionally dels it in the finally. The selftest runs ahead of every real run, so by the time _plugin_dir() is called for an actual cell the variable is gone and it falls through to the cache-dir branch.

Bit of an irony: the docstring on _plugin_dir says hardcoding one machine's path "made the ponytail/caveman arms unreproducible off that box" — and the env override that solves that is the thing getting cleared.

Fix saves the prior value and restores it. With PONYTAIL_PLUGIN_DIR=/tmp/x:

before: env after selftest = None
after:  env after selftest = '/tmp/x'

python run.py --selftest still reports all instruments valid, and the run I was after finishes now.

For what it's worth, the reproduction held up well — median src_loc on datepicker came out 401 baseline / 22.5 ponytail against your published 404 / 23, two months later on a different box.