One conformance corpus and one definitions contract for both packages

[jat255]

Fourth PR of M2. Gives the two packages a single copy of the data-dict conformance corpus and a shared fixture pinning the definitions contract they both consume. Groundwork for the definitions registry and the compiler, which land next.

The corpus moves out of pkg-r

The 14 YAML fixtures live in tests/shared/definition-export/ now instead of pkg-r/tests/testthat/fixtures/. Both suites read the same files. The alternative was a second hand-maintained copy under pkg-py/, which is the kind of drift tests/shared/ exists to prevent.

What definitions.json pins

Three things, with different provenance, which is why they are separate sections:

• export_records is generated from the data-dict binary at the pinned commit d950c5a, projected to the fields both packages consume.
• mixed_grain is not in data-dict's export at all. It is derived from the typed IR and is what call_metrics needs for its mixed-grain guard, so it is hand-maintained and the generator preserves it. One of the 42 definitions is mixed-grain, and a test asserts the values are not all identical, since an all-false fixture would pin nothing.
• invalid maps each invalid fixture to the problem code it must produce, replacing a map that was hard-coded in the R helper.

The generator refuses to run against the wrong binary

scripts/generate-definitions-fixture.sh resolves the cargo-installed binary by path rather than trusting PATH, and verifies that installation is the pinned revision. A fixture generated from another build would bless whatever that build does, in a way nothing downstream could detect.

That check took three review rounds to get right: it trusted PATH while verifying cargo, then grep -q risked killing cargo with SIGPIPE under pipefail, then || true masked cargo failing outright. Regenerating after each change produced a byte-identical fixture.

One contract detail worth knowing

data-dict omits type entirely when an expression infers no single one, such as a CASE over both a date and a datetime column. R does the same. My runner originally asserted every record has a type, which was wrong; it now pins the typeless case by name, because the obvious implementation invents a type rather than leaving it out.

Verification

284 Python tests and 219 R assertions, none skipped, so the R conformance tests genuinely ran against the pinned binary rather than skipping as they do without it. The new R assertions were checked by perturbing one translation and one grain value in the fixture and confirming the suite fails.

---

R changes, for the R reviewer

No package code changed. Everything here is under pkg-r/tests/, and pkg-r/R/ is untouched, so no exported or internal behaviour moves. Files under pkg-r/tests/testthat/fixtures/shared/ are generated by scripts/sync-shared-fixtures.sh; skip them.

What changed and why. The definition-export corpus used to live inside the R package's fixtures, where only R could reach it. Python needs the same files to check its compiler against the same cases, and a second copy would drift, so the corpus moved to the repository root and R now reads the synced copy under fixtures/shared/, exactly as it already does for the provenance and citation fixtures. Alongside that, the map from an invalid fixture to its expected data-dict problem code moved out of the helper and into the shared fixture, because it is a fact about the corpus rather than about R.

The behaviour delta. For existing tests, none: the same files are read from a different path, and every assertion that passed before passes now. What is new is that the R suite asserts the shared contract. For each valid fixture, definition_export_spec()'s contract projection and its mixed-grain result are compared against definitions.json. That means a change to the R exporter which alters a translation, an inferred kind or type, a reference list, or a grain flag now fails the R suite rather than passing silently. That is the point: the fixture claims to be authoritative, and until this it was asserted by nobody.

Blast radius. definition_fixture_paths() changed its path and is called nine times across test-definition-export.R. definition_fixture_error_code() changed its source from a literal map to the shared fixture and is called once. Six direct test_path("fixtures", "definition-export", ...) calls in test-definition-export.R and test-definition-compile.R now go through a new definition_fixture_path() helper, so the next move touches one line rather than eight. definition_fixture_contract() and definition_export_grain() are new and used only by the two new tests. Nothing outside tests/testthat/ refers to any of them.

Why it is safe, and what was checked. devtools::test_local(filter = "definition") reports 219 assertions, zero failures, zero skips. The zero-skip count matters: these tests skip themselves when the data-dict binary is absent, so a green run without it would prove nothing, and this run exercised the real binary built from the pinned commit. The two new assertions were then checked by mutation rather than by inspection, perturbing a translation string and a grain flag in the synced fixture and confirming the suite failed, then restoring it and confirming it passed.

What deserves scrutiny. The comparison sorts both sides by key. definition_export_contract() preserves authored order while the generated JSON sorts its keys for readable diffs, and I decided this fixture pins a keyed contract rather than a sequence. If you think the order definitions come back in is itself observable, that decision is wrong and the fixture should preserve authored order instead. Also worth a look is definition_fixture_contract(), which normalizes the JSON side to the R shape: it flattens each list to a character vector so that an empty sequence compares equal whether it arrived as character(0) or list(). That normalization could in principle hide a real difference in how emptiness is represented, which I judged not worth pinning.

Tests deleted. None. The only deletion is the hard-coded error-code map in helper-data-dict.R, which the shared fixture now supplies to both suites.

[github-actions]

Preview deployed to Connect (dogfood.team.pct.posit.it): https://dogfood.team.pct.posit.it/connect/#/apps/d7a36cae-8f27-448b-a478-61b81fbe3942/draft/366930

Deployed from commit 7faab7f.

[github-actions]

Preview deployed to Connect (connect.staging.pct.posit.it): https://connect.staging.pct.posit.it/connect/#/apps/ad662e1b-5048-4acc-9ad7-f9478c92274e/draft/2460

Deployed from commit 7faab7f.