Morphir example tests

Read the scenario contract and the example catalog. Use the single-file Elm scenario as the smallest working example. Keep the contract in the guide; do not create a second driver or copy MCK comparison logic. For Markdown authoring, use CLI basics. For reference Elm, follow the catalog's explicit release preparation step and select suite:elm-reference. The preparation helper stages fixtures only; repository registration, installation and compilation remain real CLI commands. Use suite:offline when no downloaded provider is available. For installed Avro/OpenAPI backends, fetch the pinned releases and run mise run examples:prepare-backends, then select suite:wasm-backends. Keep repository creation, publication and installation in scenario commands. Main backend cases compile default v4 output; backends/v3-compatibility retains explicit v3 generation coverage. Verify both when updating provider pins.

Choose the claim

State the user workflow and what passing proves. Native Elm is opt-in and currently compiles types only. Use the reference morphir-elm provider to claim function lowering. Rego assertions evaluate CLI observations; they do not establish native Morphir IR evaluation. See the evaluation architecture for the implemented provider boundary and native/WASM follow-up milestones.

Check existing tags with morphir itest examples --list. Use categorized project directories; deeper categories require no runner edit. Keep malformed source fixtures under crates/morphir/tests/fixtures/itest/.

Author and verify

1. Choose one document per example directory: scenarios.md, or a Gherkin scenarios.feature.md (or .feature) file. For a Gherkin file, put the title and tags on the Feature, the description on the scenario, and provider and workspace in a yaml itest fence in the feature description. Write commands, captures and Rego checks as the itest steps in the guide. Use @section:<id> for a stable scenario ID. For scenarios.md, use YAML frontmatter with version 1, a title, a description, tags and provider: rego. Each top-level ## heading starts an independent scenario; use ### for steps. Explicit {#id} heading attributes keep filters stable across renames. Pair yaml morphir:command, yaml morphir:assertion and yaml morphir:file metadata fences with their next language source fences. Prose, lists and subheadings may separate the pair. A pair cannot cross a scenario heading. Keep YAML outside source fences and executable fences outside list/quote containers. Unmarked fences remain documentation. Filter one heading with --filter 'category/example#id'. IDs and workspace files may repeat across independent scenarios, but must be unique within each scenario.
2. Keep ordinary source/configuration files on disk beside the document. The scenario directory is copied into an isolated workspace by default. Use workspace: {kind: "directory", path: "project", exclude: ["installed"]} in scenario metadata to choose a subdirectory or exclude custom outputs. Optional yaml morphir:file fences (or files in a yaml itest fence) add inputs; conflicting disk and inline paths fail. Use workspace: {kind: "inline"} only for a self-contained scenario that ignores adjacent project files. Commands contain literal morphir ... invocations; case names, captures and timeouts belong to their metadata fence or their steps.
3. Put complete Rego modules in assertions. Reference the most recent command by ID and enumerate unique named rule entrypoints. Only boolean true passes. Check exit status plus meaningful artifacts/diagnostics. Request stdout JSON decoding and artifact captures explicitly. Missing files, JSON null, undefined rules and errors are distinct. Never derive expected values from the implementation during a run. For exact generated text, use yaml morphir:golden with command, actual and either a paired literal source fence or a standalone expected_file. Expected file paths are relative to the scenario directory and are frozen before commands; actual paths are relative to the temporary project. Use select: {kind: lines, start: 3, end: 5} for inclusive lines or select: {kind: between, start: "BEGIN", end: "END"} for unique literal markers. Omit selection for the whole file. Exact comparison is the default; use line_endings: lf explicitly for CRLF normalization after selection. Golden checks also require exit code zero. Review fixed expected content; never regenerate it as part of the test run. Inspect the diff and retained evaluation request when diagnosing mismatches.
4. Follow root checkout setup before testing. Run mise run test:examples -- --filter <category/example> before changing the CLI. Verify independent Markdown headings, prose between paired fences and every authoring format when changing the driver. Retain real CLI subprocesses for both workflow commands and evaluation; do not replace them with internal compiler calls or an external OPA binary.
5. Diagnose with --keep-temp, inspecting command logs, observation.json and assertion request/report logs. Establish whether the source, expectation, driver or CLI is wrong before fixing it. Demonstrate that a deliberately wrong expectation fails, then restore it and rerun the selected scenario, mise run test:itest, and the offline suite.
6. Update the catalog, developer/KB guidance and Beads with the exact coverage established. Native IR evaluation, WASM hosting and installed evaluator protocol negotiation are follow-up work, not milestone 0 capabilities. Tag deliberate current-limit checks kind:negative and coverage:known-limitation, link their follow-up issue, and describe the rejection they prove. Do not count them as successful feature coverage.

Repeated tags use AND selection. Missing prerequisites fail; there are no implicit downloads or skips. Old scenario.md/test.yaml files are not executable coverage. For throwaway checks, set an outer scratch MORPHIR_HOME as well as the temporary-directory root so ordinary CLI logs stay contained. The guide documents host configuration and isolation limits.