Keep fixture annotations out of ERB blocks (#345)

[Halvanhelv]

Summary

Fixes #345.

Schema annotations were being written inside ERB tags in dynamic fixture files. For example, given a fixture that starts with a multi-line ERB block:

<%
  total = 2
%>
<% total.times do |i| %>
user_<%= i %>:
  name: User <%= i %>
<% end %>

running annotaterb models produced:

<%
  total = 2
# == Schema Information
# ...
%>
<% total.times do |i| %>
...

i.e. the annotation landed in the middle of the ERB block, corrupting the file.

Root cause

ERB fixtures cannot be parsed by Psych, so YmlParser evaluated the ERB (ERB.new(input).result) and read Psych's line numbers off the evaluated output. Those line numbers were then used as indices into the original file.

When an ERB tag spans multiple lines, the evaluated output has a different line count than the source, so the offsets no longer line up and the annotation is inserted at the wrong position — inside an ERB tag, or inside a <% ... do %> loop. Evaluating the ERB also executes arbitrary Ruby, which can fail outright.

This affected every ERB fixture, not just files starting with <% — the existing <% 1.upto(100) do |i| %> case also had its annotation written inside the loop.

Fix

Stop evaluating the ERB. Instead derive the content bounds directly from the original lines: the first non-blank, non-comment line is the start, so annotations are written above the ERB body (and after any leading comments), never inside a tag. The result is also idempotent across repeated runs.

Tests

• New YmlParser spec for a fixture that starts with a multi-line ERB block.
• New Generator specs covering an ERB fixture that starts with <% and one preceded by a comment.
• Updated the existing ERB specs in YmlParser and SingleFileAnnotator whose expectations encoded the buggy placement.

Verified these specs fail without the fix and pass with it. Full unit suite is green.