Find violations in ERB templates

[dduugg]

Fixes #1

Problem

FileCache#get_file replaced every file that isn't .rb or .rake with the placeholder string "x" before parsing. So for a violation in an ERB template, count returned 0 and every sources* method left it out, with no warning. The same happened to the other files packwerk parses as Ruby: .builder, .gemspec, .ru, Gemfile and Rakefile.

Approach

ERB templates go through Herb.extract_ruby, which blanks out everything except the Ruby code in place, so line numbers still match the template. Each tag ends with ;, so two tags on one line stay separate statements, and ERB comments and <%% escapes are dropped. The rest of the pipeline (receiver chains, anonymization, line numbers) works unchanged, and sources_with_contexts still shows lines from the template itself.

Why herb: packwerk 3.x parses ERB with better_html, which is deprecated in favor of herb, and packwerk's move away from it (Shopify/packwerk#447) uses the same Herb.extract_ruby call. Packwerk's current ERB parser couldn't be reused anyway, because it discards source positions.

Other changes:

• RUBY_FILE now matches every file type packwerk parses as Ruby, from Packwerk::Parsers::Factory.
• A file that can't be parsed now prints Could not parse <file>:<line> (<syntax error>), so its violations are not included and contributes no references. In an ERB template the line is the template line. Previously that raised a Sorbet TypeError, as did a file with no code at all (e.g. only comments). Files with no code now have no references and no warning.
• Removed the unused Violation::RUBY_FILE and its "does not play nicely with ERB files" comment.
• A constant that is a file's only code (e.g. a partial that's just <%= Billing::Invoice %>) is now found. The search skipped the root of the AST.
• sources_with_contexts shows the right lines. It used the 1-based line number as a 0-based index, so the window ran from 2 lines before to 4 after, and for a reference on line 1 or 2 the start went negative and wrapped to the end of the file. ERB partials often reference constants near the top, so this would come up a lot. Violations#sources_with_contexts also passes its start_offset: and end_offset: through now; it ignored them.
• Bumps the version to 0.1.26.

New dependency

• herb >= 0.9: 0.9.0 added the extract_ruby keyword options used here, and the new specs pass on 0.9.0 as well as 0.11.0. The options are passed explicitly rather than relying on herb's pre-1.0 defaults.
• herb is a native extension. It ships precompiled gems for macOS and Linux (arm and x86, gnu and musl), and the lockfile picks up the same platforms as nokogiri.
• Adds the tapioca-generated sorbet/rbi/gems/herb@0.11.0.rbi.

Testing

• spec/query_packwerk/file_cache_spec.rb covers references and their template line numbers in an ERB template (a block helper, two tags on one line, a comment followed by code on the same line, an escaped tag), non-HTML ERB, a constant that is a template's only code, files with no code, missing files, a syntax error (warned about once per file), each Ruby file type, and a file packwerk doesn't parse.
• spec/query_packwerk/violation_spec.rb covers count and sources_with_contexts for a violation in an ERB template, including the lines shown around each reference and a reference on the first line. violations_spec.rb covers the offsets passed to Violations#sources_with_contexts.
• 16 of the 21 new examples fail against main. The other 5 pin behavior that already worked (.rake, missing files, a template with no Ruby, Gemfile.lock). Flipping any of the three extract_ruby options, or removing the no-AST fallback, fails at least one.
• Ran against the ERB violations of a large Rails monolith: every ERB file listed in its package_todo.yml files now resolves to at least one reference (previously none), and every reported line number points at the template line containing the reported source.
• bundle exec rspec (90 examples), rubocop and srb tc pass. The new specs also pass locally on Ruby 3.3 and 3.4.