[MEDIUM] Reject executable lines in generated annotations

[OskarEichler]

Summary

• Validate the complete model annotation block immediately before reading or rewriting a target file.
• Reject any physical line that is neither blank nor a Ruby comment.
• Keep all valid default, Markdown, RDoc, YARD, and YAML annotation output unchanged.

Security impact

MEDIUM urgency. Several annotation components interpolate database schema metadata such as constraint expressions, names, index details, and enum values. If one of those values contains a newline, the following text can escape the leading # and be written as executable Ruby into a model or related source file. A later application boot or test run can execute that generated line.

Exploitation requires control of relevant database schema metadata (or direct control of the low-level annotation string) plus a developer or CI run of AnnotateRb. The gem is normally a development tool, which limits exposure, but the resulting impact is source-code execution in that environment.

Reproduction

A check-constraint component with an expression ending in a newline plus raise 'generated-code-marker' currently generates:

#  trusted_name  (price > 0)
raise 'generated-code-marker'

Passing that annotation to SingleFileAnnotator writes the uncommented line into the model. With this patch, the call raises ArgumentError before changing the file.

Verification

• Focused unsafe annotation reproduction writes an uncommented marker on main and is rejected with the target file unchanged on this branch.
• Focused valid multiline annotation is still written.
• Ruby 4.0.6 unit suite: 874 examples, 0 failures, 2 existing pending.
• StandardRB: 175 files, no offenses.
• All 98 lib Ruby files compile.
• git diff --check.

Limitations

This validates model and related-file annotations at their final write boundary. Route annotation generation uses a separate path and prefixes every route-map row itself; configuration files and the project Rakefile remain trusted local code.

Breaking changes

Callers that intentionally pass non-comment content through the low-level model annotation API now receive ArgumentError. Generated annotation blocks are documented and expected to contain comments only.