Add a `rails` policy to the policy registry

[bkd-dotcom]

What this is

signetry-core ships a policy registry: a directory of YAML files, one per repository
shape, that says which paths an AI agent may touch for a given task and which it may not.
Six ship today (signetry_core/policies/):
python-library, node-service, monorepo-service, docs-only, dependency-bump,
ci-workflow-fix. A Rails application is missing.

This is the highest-leverage thing you can contribute without touching the kernel: one new
file, no Python, and no test to write — tests/test_policy_registry.py is parametrized
over every file in that directory, so your policy is validated the moment it exists.

The deliverable

One file: signetry_core/policies/rails.yaml. Copy
python-library.yaml
as your starting shape — the # @policy header comments are load-bearing metadata, not
decoration.

A starting point, not a spec to type in verbatim — argue with it, the reasoning is the
contribution:

# @policy id: rails
# @policy title: Rails app (app/ + spec/, RSpec)
# @policy summary: A Rails application. The agent may change models, controllers, views, jobs and specs,
#   and must keep RSpec and RuboCop green. Migrations, the generated schema, initializers
#   and credentials stay off-limits.
# @policy stack: ruby, rails, rspec, rubocop, activerecord
# @policy author: your-github-handle
# @policy blocks: db/migrate/20260901120000_add_plan.rb, db/schema.rb, config/initializers/cors.rb, config/credentials.yml.enc, .github/workflows/deploy.yml
# @policy allows: app/models/subscription.rb, app/controllers/billing_controller.rb, spec/models/subscription_spec.rb, config/routes.rb, Gemfile
version: 2
task_type: feature-work

allowed_paths:
  - "app/**"
  - "spec/**"
  - "test/**"
  - "lib/**"
  - "config/routes.rb"
  - "config/locales/**"
  - "Gemfile"
  - "README.md"
  - "CHANGELOG.md"

forbidden_paths:
  # db/schema.rb is GENERATED by running migrations. Hand-editing it makes the
  # checked-in schema disagree with the migration history, and the next migration
  # run silently overwrites the edit.
  - "db/**"
  # Inside config/** — routes.rb and locales are allowed above, but initializers
  # run at boot before any spec loads, and can monkey-patch anything in the tree.
  - "config/initializers/**"
  - "config/credentials*"
  - "config/master.key"
  - "config/environments/production.rb"
  - ".github/**"
  - "Dockerfile*"
  - "**/.env*"
  - "**/*secret*"

max_files_changed: 15
required_checks:
  - "bundle exec rubocop"
  - "bundle exec rspec"

policy_owner: your-team
policy_version: "1.0"

The part people get wrong

Your blocks list must include at least one path that sits inside your own
allowed_paths. Carving an exception out of a directory you otherwise own is the whole
skill. For a Rails application that exception is:

config/initializers/** — you want config/routes.rb and config/locales/**
allowed, so the natural instinct is to allow config/**; do that and the agent can
monkey-patch the framework from a file that runs before any spec.

The Rails trap worth a comment in the file

Gemfile without Gemfile.lock (or the reverse) is a build that resolves
differently on the next machine. And Rails' autoloading means adding a file under
app/ changes behaviour without any call site appearing in the diff — a new
app/models/concerns/ module gets picked up by name.

Acceptance criteria

• signetry_core/policies/rails.yaml exists; filename matches @policy id.
• blocks and allows each list 3–4 realistic paths for this stack, and don't overlap.
• At least one blocks entry is inside allowed_paths.
• Every forbidden_paths entry that isn't self-evident carries a comment saying why,
  not just that it is. python-library forbids conftest.py at any depth and the comment
  explains it executes at collection time on every developer machine — that is the bar.
• pytest tests/test_policy_registry.py is green. The suite proves each claimed block is
  actually refused by the real evaluate_contract, so a policy that misleads adopters
  fails CI rather than shipping.
• If the policy permits something risky, it says so in a caution: — signetry init
  prints it at adoption time. A risky policy with no caution gets sent back; a risky
  policy that's honest is fine.

Getting started

The full numbered walkthrough, including what makes a policy worth merging, is in
docs/site/policy-registry.md → "Contributing a policy".

git clone https://github.com/Signetry/core && cd core
uv sync
uv run pytest tests/test_policy_registry.py -q
uv run signetry policies          # your entry should appear here once the file exists

Comment to claim it — one policy per contributor so nobody's work gets duplicated. Happy to
review a half-finished scope list; the reasoning matters more than the YAML.