S7 (#188) — one declared way to find views-datafactory; Cluster K closes

[Polichinel]

Closes #188. Closes register C-46. Closes Cluster K. Part of epic #181.

What C-46 actually cost

Four sites resolved views-datafactory four different ways. One was:

_DF = Path("/home/simon/Documents/scripts/views_platform/views-datafactory")

So the repo's only cross-repo release gate had never run anywhere but one laptop. "Skipped in CI" undersold it.

site	was
test_datafactory_deploy_readiness.py	an absolute path to one developer's home directory
test_gaul_lookup_fidelity.py	its own _REPO.parent / "views-datafactory"
test_delivery_coverage.py	its own parents[2] / … — a fourth site the issue did not list
scripts/build_gaul_lookup.py	the shape that was already right

Now: grep -rn "/home/\|/Users/\|C:\\" over tests/, scripts/ and the package → 0. With the sibling present the deploy gate runs (1 passed, 3 xfailed, its deliberately-tuned markers untouched). Without it: clean skips, no collection errors. Every skip names VIEWS_DATAFACTORY and the conventional path.

Two resolvers, on purpose

A script must not import from tests/ — dependency direction — and the contracts differ: the script returns a Path even when absent so main can raise naming both the flag and the variable; the helper returns None because a missing sibling is a normal skip.

WET before DRY. What is guarded is not the duplication but the agreement, because a builder writing from one checkout while the tests verify against another would report success on a lookup compared to a producer it was not built from.

Review sharpened that guard, and it was the right catch

The agreement test first skipped when no checkout was present. But both resolvers read $VIEWS_DATAFACTORY first, so they can only diverge in the fallback — which is exercised precisely when nothing exists there. The one case the test was for was the case it declined to check, and it declined in CI, where it matters most.

Fallback paths are now compared unconditionally. Mutation-proven:

E  AssertionError: the builder's fallback and the tests' fallback resolve different
   directories; with no environment override they would disagree silently

Review also caught this change contradicting itself

• Cluster K's Resolution scope: line still called C-60 and C-46 "partial" — eight lines above the note closing them.
• C-46's body still asserted a [backlog] state that ended on resolution.

Both fixed. A cluster that closes with a contradiction inside its own block undercuts the closure, and this epic exists to stop exactly that.

Cluster K closes — five entries, all full

C-59, C-61 (build-time guarantees) · C-43 (the value-correctness debt) · C-60 (the declared lookup_version) · C-46 (the machine-specific path).

The fix strategy predicted the shape correctly. What it did not anticipate is recorded in the closing note:

A cluster about "the artifact is trusted but unverified" had a verification apparatus that was itself partly unverifiable — the gated half it prescribed could not run for anyone but the maintainer, and the provenance stamp it relied on could silently become the string "unknown".

What remains is not this repo's: whether views-datafactory's area-majority join is right at high latitudes is #387 there. C-43's scope split holds.

The CI question — recorded, not implemented

The issue put it out of scope, so this states a recommendation with reasoning and a named trigger rather than acting:

Do not add sibling checkouts to the per-PR workflow. It couples this repo's CI to another repo's default branch — an unrelated upstream commit turns this repo red, which is the flapping TestReleaseGate already documents and was re-pinned for once. The always-on halves guard the committed artifact; the gated halves answer "is the producer still consistent with us?", a scheduled question, not a merge gate. If wanted: a weekly workflow that opens an issue on divergence.
Trigger: the next time an upstream change reaches FAO through this repo without anyone noticing first.

Three gated groups now share that decision — this gate, the fidelity suite's producer half, and C-57's registry half. Flagged for S9 (#190) so it lands as a decision someone takes rather than a note someone finds.

Verification

ruff check .                                                       # clean
pytest -q                                                          # 299 passed, 3 skipped, 41 xfailed, 5 known pyarrow
VIEWS_DATAFACTORY=/nonexistent pytest -q                           # 293 passed, 12 skipped — no errors
grep -rn "/home/\|/Users/" tests/ scripts/ views_postprocessing/   # 0
git diff --stat origin/main -- tests/fixtures/                     # empty

Register: 74 concerns, 18 open, 56 resolved.

🤖 Generated with Claude Code