DR-117: Pipeline fixtures live with the pipeline, not with the tests

DR-117: Pipeline fixtures live with the pipeline, not with the tests

AMENDED by DR-142 (2026-08-10). This ruling’s filing rule stands and is what decides where the Python tests live. Two things below are corrected: the “anything under scripts/cocoindex_pipeline/ moves as one unit” premise is false at the directory boundary (six measured cross-boundary edges; onprem-deploy.yml packages scripts/ whole), and the deferred destination is now settled as a singular root-level pipeline/ holding the runtime bundle plus its tests — which is the open question this record routed forward.

Context

The form-extraction template fixtures — real product inputs, including a British Council RFP and a standard selection questionnaire — have moved twice without a ruling, and each move was made to fix the symptom the previous location caused.

They began at docs/testing/test-data/templates/. That tree is outside the sidecar’s pruned build context (cp -R scripts + requirements.lock), so verify_driver.py running in-container could not read them, and the nightly compensated with a docker cp. At S521–S522 they were moved to scripts/tests/fixtures/form-extraction/templates/ (b9c55abf, 90cc3b03, eb9ccac9), which fixed the build-context problem.

The owner’s ruling at S522 is that neither location was ever correct, and that the second move is what surfaced the real defect: the vendored synthetic Platform corpus at scripts/cocoindex_pipeline/fixtures/platform-corpus/ had been missed entirely as the walked baseline (id-412). With that in view, the question “where do the form fixtures live” stops being a tidiness question and becomes a question about which axis fixture trees are filed on.

Two ratified constraints frame the answer, and they pull in opposite directions if read naively. D3 (RATIFIED S511, the owner having explicitly overruled a proposal to retire the two orphan fixtures) says they should “sit with the other corpus forms”, with the exact destination deferred until the manifest lands — that is now, via id-406. DR-014 retires the corpus forms-route, and platform-corpus-shape.test.ts enforces it: the tree must contain exactly ten entries, with forms/ and __qa__/ named as prohibited.

A third constraint arrived from an unrelated direction. DR-116 settled that palace rooms key on directory name matched at any depth, and the S521 rooms review left an open finding: 99 Python pipeline tests route to room tests rather than pipeline, because tests is a token of __tests__ and no part of scripts/tests/** says cocoindex. Anything filed under scripts/tests/ is therefore roomed by kind, not by domain — the opposite of the ratified intent.

Decision

Fixture trees are filed by the domain that owns them, never by the kind of artefact they are. The form-extraction templates settle at:

scripts/cocoindex_pipeline/fixtures/form-templates/

— a sibling of platform-corpus/ under the pipeline’s own fixtures root, not a child of it and not under scripts/tests/.

This resolves D3 and DR-014 together rather than at each other’s expense: the fixtures root is “with the other corpus forms”, while platform-corpus/ keeps its exactly-ten-entries shape and its no forms/ invariant untouched. It also extends an already-ratified anchor rather than inventing a location — deploy/onprem/sync-platform-corpus.sh:8 names scripts/cocoindex_pipeline/fixtures/platform-corpus/ as “RATIFY-1: the single source of truth”.

Two consequences follow that a purely local reading would miss. Under DR-116 these files now room by domain, because cocoindex_pipeline carries the matching token — which is also the shape of the answer owed to the S521 scripts/tests/ finding. And when cocoindex moves out of scripts/ (agreed in principle, deferred), anything under scripts/cocoindex_pipeline/ moves as one unit, while anything under scripts/tests/ would be orphaned by that move.

Alternatives Considered

• Keep them at scripts/tests/fixtures/form-extraction/. Rejected. It files a pipeline artefact by kind, it rooms to tests under DR-116, and the deferred cocoindex relocation would strand it — guaranteeing a third move of the same files. It is also the location the owner has ruled is not the correct approach.
• Move them inside platform-corpus/, e.g. as platform-corpus/forms/. Rejected on a hard constraint, not a preference: DR-014 retires the corpus forms-route and platform-corpus-shape.test.ts asserts exactly ten entries with forms/ prohibited. Adopting this would require weakening a ratified guard to accommodate a filing choice.
• Return them to docs/testing/test-data/. Rejected. Outside the pruned build context, so the in-container driver cannot read them without the compensating docker cp that caused the original problem.
• Leave the question open until the sidecar leaves scripts/. Rejected. id-406’s manifest is landing now and must record a destination; deferring means the register whose entire purpose is to replace hardcoded paths would itself hardcode a known-wrong one.

Consequences

• id-406 settles AC-7 with this destination and registers the tree with staging_mode: per-test, against platform-corpus/’s walked baseline.
• id-412 performs the move and repoints all 24 consumers — 21 TS integration tests (through the new shared fixtures module), .github/workflows/cocoindex-nightly.yml, scripts/cocoindex_pipeline/verify_driver.py, and scripts/tests/test_form_extractors.py.
• The rule generalises beyond these files. Any future fixture tree is filed under the domain directory that owns it. __tests__/fixtures/** remains correct for app-side fixtures, because the app’s tests are the owning domain there.
• A partial answer to the S521 scripts/tests/ room finding. It does not settle where the 99 Python pipeline tests should live — that stays open and belongs with the deferred cocoindex relocation — but it removes the fixture trees from the disputed ground.
• Three defects surfaced while deriving this, each routed rather than fixed here: the nightly’s docker cp of the form fixtures is likely redundant now that they ship in the build context (id-412); the workflow’s claim that the image “never carries … tests” became false when they moved under scripts/ (id-412); and five files still cite dead docs/testing/ paths, including an operator-run IP-leak sweep that now silently finds zero files where it expects eight, on a public repo (id-128 W9).