Cookbook fixtures

The cookbook is one tree with two halves, and every recipe in either is a fixture run by CI with the same commands a user runs. A recipe that is not run is prose, and prose rules do not hold (plan 0005 § 1.1).

Folder	Holds	Runner	Plan
`guards/<slug>/`	the guard catalogue: one recipe per guard, over a committed graph	`guards/run.sh`	plan 0005
`questions/<slug>/`	one page per architecture question, over a small source repository the docs site's playground loads	`questions/run.sh`	plan 0004, Step 5

Both runners run in the guards CI job; the docs site renders the question pages from their fixtures (ADR-0074). The rest of this page is the guard catalogue; questions/README.md describes a question fixture.

Plan: plan 0005, Step 1
Requirements: FR-RULE-01, FR-CFG-07, FR-REACH-01 (the question pages)
Decisions: ADR-0004 (the graph shape), ADR-0005 (what conversion drops), ADR-0007 (no vacuous recipe), ADR-0021 (public commands only), ADR-0024 (a regenerated expectation is explained), ADR-0074 (the question pages)

A fixture

guards/<slug>/, where the slug is kebab-case and names the guard, not the rule:

File	Required	Holds
`rulebearing.yaml`	yes	the recipe: each rule with `comment` (ending in a decision token), `fix`, `severity` and `examples` with at least one `forbidden` and one `allowed` edge
`graph.json`	yes	a small graph in dependency-cruiser's `cruise-result` shape, holding only the modules the recipe needs, with every module fact as the extractor writes it
`expected.json`	yes	`summary.violations` and `summary.vacuousRules` exactly as `cruise -T json` prints them
`README.md`	yes	what the graph holds and why, which extractor supplies each fact in a real run, and any difference from the cookbook export
`expanded.yaml`	when the recipe uses `defines` or a shorthand	`config expand`'s output
`converted/.dependency-cruiser.json`, `converted/dropped.txt`	for a dependency-layer recipe	`config convert --to dependency-cruiser`'s output and its report of what it dropped
`budget.json`	for a ratchet recipe	the budget file
`negative/<case>/rulebearing.yaml`, `exit-code`, optional `graph.json`	when the recipe has a failure to prove	a configuration that must stop `cruise` with that exit code (ADR-0008)

The graph makes every rule non-vacuous, so vacuousRules is empty, except in a recipe whose point is allowEmpty.

Running

cargo build --release
cookbook/guards/run.sh no-circular     # one fixture
cookbook/guards/run.sh --all           # every fixture, one line each

guards/run.sh lists the commands it runs. RB names another binary.

Adding a recipe

Write rulebearing.yaml, graph.json and README.md.
Run RB_UPDATE_SNAPSHOTS=1 cookbook/guards/run.sh <slug> to write expected.json, expanded.yaml and converted/.
Read what it wrote. The violations must be the ones the README says the graph holds, and nothing else.
Run cookbook/guards/run.sh <slug> without the variable.

A recipe that does not behave as the cookbook says is a finding. Correct the recipe in the fixture and say so in its README; the dated cookbook export in docs/artifacts/ is not edited. If the engine is wrong rather than the recipe, record the defect in plan 0005's status table against the plan that owns the engine.