Scan for committed secrets in CI: non-public Babel URLs and Google Sheet IDs

[gaurav]

Why

"Never commit the Translator-specific URL" has been a documented rule in CLAUDE.md and
README.md for as long as there has been a config template, and until #20 nothing enforced it. The
rule had already failed in practice: the internal releases URL was the hardcoded default in
cli.py from the initial commit until the history was rewritten to remove it on 2026-09-01.

#20 added TestCommittedConfigTemplate (tests/test_cli.py), which fails if a non-public host
appears in env.default. That closes the template, but not the hole the leak actually came
through — a default value in source. A scan over the whole tree is what would have caught it.

What to add

A CI step (and ideally a test, so it also runs locally) that fails on:

1. Non-public Babel URLs. Every stars.renci.org URL in a tracked file must be under
   /var/babel/. Today there are 11, all compliant:

   6  https://stars.renci.org/var/babel/
   5  https://stars.renci.org/var/babel/latest/
   

   Prefer an allowlist ("only these paths are permitted") over a denylist naming the private
   directory — writing the private path into ci.yml to grep for it would reintroduce the string
   this repository just spent a history rewrite removing.

2. Google Sheet IDs. Requested for coverage; note there is currently no Google Sheets
   reference anywhere in this repo, tracked or in history (git log -S docs.google --all is
   empty), so this one is preventive rather than fixing a known leak. Relevant because upstream
   Babel curates some concords in Sheets, so an ID could plausibly arrive here with a future
   feature. Pattern: docs.google.com/spreadsheets/d/<44-char id>, plus bare 44-character
   [A-Za-z0-9_-] IDs assigned to a variable named like a sheet.

Notes

• Run it over tracked files only (git grep), not the working tree, so a developer's local .env
  is out of scope — .gitignore covers .env and .env.* as of Add babel-explorer: a CLI for querying Babel cross-references via DuckDB and NodeNorm #20.
• Consider a general secret scanner (gitleaks, trufflehog) instead of bespoke greps, if the
  maintenance trade seems right. A bespoke allowlist is easier to reason about and produces no
  false positives; a scanner catches classes nobody thought to write a rule for.
• This does not retroactively fix anything: the URL is out of this repository's history, but
  GitHub still retains pre-rewrite objects reachable by SHA, and the tree it pointed at still
  serves 200 unauthenticated. Both are being handled outside this repo.