E2E test of the manager orchestrating the delivery (C-03's residual) — BLOCKED on a non-production Appwrite project

[Polichinel]

Rewritten 2026-08-02. The original body described the pre-#149 manager and is preserved at the bottom. This issue is now register C-03's sole home — C-03 closed in #200 with its manager-orchestration residual relocated here by number.

What is actually missing

Every leg of the delivery is covered:

leg	tests
enrichment	test_enrichment.py (16)
artifact build	test_historical_builder.py (7)
reader parity	test_historical_parity.py (3)
invariants on primitives	test_input_integrity_e2e.py (8)
the ADR-013 wire, end to end	test_hop_b_sink_e2e.py (6)
the metadata null-gate	test_validation.py (14)

What is not covered is the manager orchestrating them — _read → _transform → _validate → _save through a real UNFAOPostProcessorManager, with pipeline-core's Template Method actually driving the lifecycle.

That gap is real. Every test above verifies a piece in isolation or a seam in a subprocess; none proves the pieces are wired together in the order the base class calls them.

Why it is blocked, precisely

Not on effort. On an environment that does not exist.

1. views-pipeline-core must be installed — true, and no longer the hard part. contract/ and delivery/ are dependency-light and import without it (test_clone_readiness.py proves this in a subprocess). Only the manager needs it.
2. A production-like Appwrite project must exist — and it does not. þing-02 D2 forbids integration tests against the production project, and no non-production project has been created. Mocking DatastoreModule deeply enough to exercise _save_contract's manifest-last commit ordering would mean reimplementing the store, at which point the test proves the mock rather than the delivery.

Creating a non-production Appwrite project is an operator decision — it costs money and console access (CLAUDE.md). It is not something this repo can decide its way around.

Named trigger

Per ADR-014 §4, a deferral needs a trigger and an owner rather than "later":

Trigger: a non-production Appwrite project exists, or þing-02 D2 is amended.
Owner: the operator, since both are operator decisions.

Until one of those happens this issue should not be picked up, and picking it up would produce a test of a mock.

What NOT to do

• Do not mock your way to a green e2e test. A test that stands in for the store proves the stand-in. The wire's guarantees — content hashes, header asserts, §6 no-collapse, gid parity, manifest-last — are already covered by test_hop_b_sink_e2e.py against real bytes.
• Do not add a redundant test to close a register entry. C-03 is already closed; this issue is not blocking anything, and inventing coverage to tick it would be worse than the gap.

Standing marker

tests/test_falsification_campaign_3_5.py remains the in-suite marker. Note its limitation, recorded in C-36: it is a pure assert False probe, so it cannot flip to XPASS when the gap closes. Whoever closes this must retire that probe by hand — S4 (#200) set the precedent, removing three discharged probes with a dated DISCHARGED note in the module docstring.

Stale requirements from the original body, corrected

• "Verify upload metadata includes the enrichment_description timestamp" — _delivery_description and _append_metadata were deleted in S1 — Delete the retired delivery path; refuse an incomplete config #149. Provenance is now built by delivery/provenance.py and reaches the store through _historical_frame_description.
• "Verify the output parquet has all 9 metadata columns, non-null" — _validate no longer null-gates. That moved to contract/historical.assert_metadata_complete, firing at artifact-build time, and is covered by test_validation.py.

---

Original body (2026-06-04), preserved

Context

The falsification campaign (Claim 3.5, FALSIFIED) confirmed that no end-to-end pipeline test exists. The full pipeline (read → transform → validate → save) is never tested as an integrated unit.

Blocker

views-pipeline-core is not installed in the test environment. UNFAOPostProcessorManager inherits from PostprocessorManager and ForecastingModelManager — both require views-pipeline-core.

Requirements

When views-pipeline-core is available:

1. Create a test that mocks ViewsDataLoader and Appwrite DatastoreModule
2. Exercise the full pipeline: _read() → _transform() → _validate() → _save()
3. Verify the output parquet has all 9 metadata columns, non-null
4. Verify upload metadata includes the enrichment_description timestamp

Risk Register

C-03 (Tier 3, test coverage gaps). Campaign Claim 3.5.
Test stub: tests/test_falsification_campaign_3_5.py

[Polichinel]

Status vs current development (2026-06-27): partial coverage now exists; full real-manager test still blocked.

What now exists (replica pattern, since the manager can't be instantiated without pipeline-core — C-03):

• tests/test_input_integrity_e2e.py (S6/S6 — End-to-end input-integrity test suite (through the manager seam) #57) replicates the manager's extract→invariant chain end-to-end across all five guards.
• tests/test_validation.py (_validate) and tests/test_append_metadata.py (_append_metadata) replica-cover those stages.

Still missing (the actual ask): a single test exercising the real _read()→_transform()→_validate()→_save() on the live UNFAOPostProcessorManager with mocked ViewsDataLoader + DatastoreModule. Blocker unchanged: pipeline-core not importable in the vpp test env (C-03).
Note: the un_fao smoke test (views-models, 2026-06-26 postmortem) exercised the real manager end-to-end against live infra as a manual proxy — verifying _read_historical_data+_append_metadata on 5.7M real actuals — but that is manual, not a CI test. Issue stays open, scope narrowed to the real-manager CI test.

[Polichinel]

Stale requirements — this describes the pre-#149 manager.

Two of the four requirements name things that no longer exist:

• "Verify upload metadata includes the enrichment_description timestamp" — _delivery_description and _append_metadata were deleted in S1 — Delete the retired delivery path; refuse an incomplete config #149. Zero hits in the package.
• "Verify the output parquet has all 9 metadata columns, non-null" — _validate no longer null-gates. That moved to contract/historical.assert_metadata_complete, which fires at artifact build time.

The stated blocker also softened: the delivery machinery in contract/ and delivery/ is now dependency-light and imports without views-pipeline-core (proven by tests/test_clone_readiness.py). Only the manager itself still needs it.

This issue is C-03's second residual — the missing enrich→validate end-to-end path — and epic #181's S4 (#185) is scoped to either close it or state in writing why it stays open. S4 will check whether test_historical_parity.py and test_historical_builder.py already cover it before anything new is written; adding a redundant test to close an entry is not the goal.

[Polichinel]

This issue is now C-03's sole home in this repo. C-03 was closed in #200 (S4 / epic #181) with its manager-orchestration residual relocated here by number.

Body requirements are stale — see my earlier comment; enrichment_description and the 9-column _validate null-gate both describe the pre-#149 manager.

What is actually still missing, as of 2026-08-02. Every leg has coverage:

leg	tests
enrichment	test_enrichment.py (16)
artifact build	test_historical_builder.py (7)
reader parity	test_historical_parity.py (3)
invariants on primitives	test_input_integrity_e2e.py (8)
the wire, end to end	test_hop_b_sink_e2e.py (6)
the metadata null-gate	test_validation.py (14, new)

What is uncovered is the manager orchestrating them — _read → _transform → _validate → _save through the real UNFAOPostProcessorManager. That needs views-pipeline-core installed and a production-like Appwrite environment, and þing-02 D2 forbids integration tests against the production project (no non-production project exists). So this is blocked on an environment decision, not on test-writing effort.

tests/test_falsification_campaign_3_5.py remains the standing marker.

[Polichinel]

Rewritten today. Three things changed and one did not.

Changed: this is now C-03's sole home (C-03 closed in #200, residual relocated here by number); the body's stale requirements are corrected — enrichment_description and the 9-column _validate null-gate both describe the pre-#149 manager; and the issue is labelled blocked with a named trigger per ADR-014 §4, rather than sitting unlabelled as if it were merely unstarted.

Did not change: the gap is real. Every leg of the delivery is covered; the manager orchestrating them is not.

The blocker is an environment, not effort: þing-02 D2 forbids integration tests against the production Appwrite project, and no non-production project exists. Creating one is an operator decision — money and console access. Trigger: a non-production project exists, or D2 is amended.

Also recorded: whoever eventually closes this must retire tests/test_falsification_campaign_3_5.py by hand. It is a pure assert False probe and cannot flip to XPASS when the gap closes — C-36's known residual. S4 (#200) set the precedent for discharging one.