Fix annotation placement in nested files and around trailing user comments

[yamat47]

Problem

Two real-world layouts make annotaterb misplace or corrupt the schema annotation.

1. The end of an existing annotation gets mis-detected

AnnotationFinder walks backward looking for the first # separator. With a multi-paragraph user doc placed after the schema, the walk stops inside the user's doc:

# == Schema Information
# Table name: users
# ...
#                                  # ← actual end of schema
# First paragraph of user doc.
#                                  # ← walk-back stops HERE (wrong)
# Second paragraph of user doc.
class User < ApplicationRecord

The first paragraph and the inner # are absorbed into the detected annotation range and silently lost when the annotation is rewritten.

The same logic fails when the annotation lacks its trailing # separator (e.g. manually trimmed): the walk stops at a section-internal #, and the trailing content row escapes the detected range — duplicating on the next run.

2. The annotation target line is wrong in nested files

Generator#determine_annotation_position previously used parsed.starts.first (no class case) and parsed.starts.select { |name, _| type_map[name] == :class }.last (class case). Both fail on common shapes:

# spec file: parsed.starts.first picks `require`,
# so the annotation lands above the require line:
# == Schema Information       # ← wrong: should be just above RSpec.describe
# ...
require 'rails_helper'

RSpec.describe Sample do
end

# self-reference inside the method body: `.last` picks
# the body reference, so the annotation lands inside the method:
class Sample < ApplicationRecord
  def call
    # == Schema Information    # ← wrong: should be above `class Sample`
    # ...
    Sample::Helper.run
  end
end

Solution

Two small post-processors over CustomParser's raw output.

walk_forward_through_schema (in AnnotationFinder, replaces the walk-back). Advances from the schema header while every successor matches schema_like? — a header prefix, an exact header (Indexes / Foreign Keys / Check Constraints), a tabular row (# + 2+ leading spaces), or a # separator. Stops at the last schema-like comment.

FileParser::AnnotationTarget.find (new). Skips require / require_relative / load and de-duplicates class entries by name so each class only contributes its declaration line. Generator and ParsedFile both go through it.

[drwl]

Thanks for the write up and making these changes!