The new-screenshot error prints a path that will never exist in CI

[pftg]

Under ADR-010 (2.0 warns, 2.1 deletes) this is a 2.0 blocker — it is a live bug regardless of the release plan.

The bug

lib/snap_diff/screenshot_matcher.rb raises, on a screenshot with no baseline:

No existing screenshot found for <path>.png!
To record it: run the test, then `git add <path>.png` and commit -- baselines are read from git.

Under CI that file is never written. build_screenshot_assertion calls check_base_screenshot before capture_screenshot (screenshot_matcher.rb:24-31), and fail_if_new defaults to !ENV["CI"].nil? (config.rb:106) — so in CI the raise happens first and nothing lands on disk.

$ CI=true bin/rails test test/system/about_page_test.rb
Failure: No existing screenshot found for .../doc/screenshots/about_page.png!
$ ls doc/screenshots/about_page.png
ls: no such file
$ CI= bin/rails test test/system/about_page_test.rb
$ ls doc/screenshots/about_page.png
doc/screenshots/about_page.png     # written

The instruction is unfollowable on the exact run that prints it. Two independent judges on the #259 panel found this ordering defect without being told about it.

This is the rule from #259 broken in a message we wrote ourselves: never print a command in an error message that is not generated from live state. Same shape as RECORD_SCREENSHOTS=1, which sat in this message for years pointing at a feature that never existed.

Checklist

• Reproduce first — failing test before any fix
• Capture before the raise, so the file exists whenever we name it. Careful: check_base_screenshot also performs the git checkout that drives need_to_compare? — only the raise moves, not the checkout
• Message becomes a function of live state: name the path only when the file was actually written
• Guard test written as the USER's code, not an internal unit test
• Mutation-check: revert the fix, confirm the test reds, restore (targeted Edit, never git checkout <file>), confirm green
• rake test:unit + rake test:canonical + standardrb lib test green
• Blocks: the accept workflow in Accept-workflow and honest reporting: adopt what Playwright, VCR and SimpleCov already proved #259 — flipping any default without this gives failure without the file

Related, same file, not yet fixed

• lib/snap_diff/errors.rb:12-14 asserts "EVERY error this gem raises inherits SnapDiff::Error". True at the raise sites, misleading about what a user sees: integrations/minitest.rb:38-39 converts ExpectationNotMet to Minitest::Assertion, integrations/rspec.rb:60-61 to RSpec::Expectations::ExpectationNotMetError. The docs were corrected in docs: stop teaching two commands that do not work #257; the code comment now contradicts them
• SnapDiff::ErrorWithFilteredBacktrace does not inherit SnapDiff::Error — it is its superclass