5 pre-existing test failures surfaced by the first full cold-cache (--rebuild-caches) run

[ligon]

The 2026-06-14 WB-parity loop ran a full cold pytest --rebuild-caches gate (1878 tests) — the first such full cold run in a while. It found 6 failures: 1 was introduced by the parity work and is fixed (#444, Mali u-code leak); the other 5 are pre-existing and were latent behind warm caches / scoped per-gap gates. git diff 7bdd03f2..development confirms none of these features/countries were touched this session.

The 5 pre-existing failures

1. test_sample.py::test_covers_all_waves[Albania] — Albania sample missing waves: ['1996']. Pre-existing sample-coverage gap (same class as the EthiopiaRHS one fixed in PR feat(EthiopiaRHS): canonicalize structure + promote 2004/2009 to full waves #429). Fix: wire Albania's 1996 wave into sample() (or document the exclusion).

2. test_uganda_invariance.py::test_parquet_matches_baseline[2005-06/_/interview_date.parquet] — column mismatch: actual ['Int_t'], expected ['date']. The interview_date feature output evolved (date→Int_t); the stored baseline is stale.

3. test_uganda_invariance.py::test_parquet_matches_baseline[var/cluster_features.parquet] — dtype mismatch: Latitude/Longitude actual float64, expected float32. Stale baseline dtype.

4. test_uganda_invariance.py::test_parquet_matches_baseline[var/shocks.parquet] — baseline mismatch.

5. test_uganda_api_vs_replication.py::test_api_matches_replication[interview_date] — no shared data columns after merge — same root cause as Add the second half Uganda interview_date.py files #2 (interview_date date→Int_t).

Nature / fix

These are stale-baseline drift (2–5: the Uganda baseline parquets predate legitimate feature-output evolution) and one coverage gap (1: Albania 1996). Resolution is a maintainer call: regenerate the Uganda baselines (after confirming the current output is correct), and fix/justify the Albania 1996 coverage. Not introduced by the parity merges, but they make a full cold CI red — worth clearing before the pending version bump / master merge.

Surfaced by the release-gate run logged in slurm_logs/2026-06-14_parity_loop/release_gate_fulltest.out.

🤖 Generated with Claude Code

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

[ligon]

Second full cold gate (post Mali-fix #444 + #443, -n 24) — confirms the parity work is clean and adds a flakiness finding.

5 failures (down from 6 — the Mali u-leak is fixed). All 5 are the Uganda-baseline / Albania cluster; none attributable to the parity merges. Importantly, the failing set is not stable across runs:

• Stable both runs: Albania covers_all_waves[1996], Uganda cluster_features (dtype), shocks baseline, api interview_date.
• Shifted: gate 1 failed interview_date.parquet baseline (passed gate 2); gate 2 failed api food_quantities (passed gate 1).

food_quantities[Uganda] passed gate 1, which already contained all parity merges — so our work didn't break it. The shift points to cross-test cache-state dependence under parallel --rebuild-caches (these baseline/replication tests read var/*.parquet built by other tests; xdist worker distribution changes whether the parquet is present at read time).

So the fix for this issue is two-pronged: (a) regenerate the stale Uganda baselines + fix Albania 1996 (the real content issues); (b) make the Uganda invariance / api-vs-replication tests robust to parallel cold runs (e.g. build their inputs in-fixture, or mark serial), so a full cold CI is deterministic.

Gate log: slurm_logs/2026-06-14_parity_loop/release_gate_FINAL.out.

[ligon]

Investigated all 5; resolved 4 in #447, recharacterized the 5th. None were introduced by the parity loop.

Fixed in #447:

1. test_sample[Albania] "missing 1996" — framework footgun: Country.waves reads a _/{country}.py module's waves attr before the data_scheme Waves: list, then falls through to a directory glob when the module defines none. albania.py had no waves, so the glob re-included 1996/ (has Documentation/SOURCE.org) despite data_scheme documenting it excluded (EWS, not LSMS). Fixed by adding the authoritative waves to albania.py. (Albania is the only country currently hitting this; the broader property fix — make data_scheme Waves: authoritative even when a module exists — is left as a framework follow-up.)
   2–4. Uganda invariance[interview_date / cluster_features / shocks] — stale baselines tracking intentional, shipped changes (interview_date date→Int_t enhancement: interview_date schema cleanup (missing v level; Int_t as string; old declarers leak v/date columns) #325; cluster_features GPS wiring float32→float64 Uganda cluster_features: wrong grain, stringified-float District, missing GPS wiring #161; shocks pandas-3.0 dtype migration). Baseline regenerated from a fresh cold build, restricted to the original 53-key set; all three content_hash changes verified deterministic across two cold builds. Validated 53/53 (--no-purge).
2. api_vs_replication[interview_date] — replication baseline predates enhancement: interview_date schema cleanup (missing v level; Int_t as string; old declarers leak v/date columns) #325; xfail'd to match the existing food_acquired/nutrition pattern.

Recharacterized, NOT a stale baseline:

• api_vs_replication[food_quantities] passes deterministically warm AND cold in isolation. Its gate failure is the pre-existing parallel-cold cache-race flake → test_table_structure flakes under parallel xdist (collection-time enumeration vs concurrent cache mutation) #330, not a baseline issue. An xfail would XPASS, so it's left alone.

Post-fix full cold gate (-n 24 --dist=loadfile --rebuild-caches, 40 countries, f840977c): 1117 passed, 137 skipped, 11 xfailed, 1 xpassed, **0 failed** (was 5 failed).

Suggest closing this once #447 merges; the residual parallel-cold flakiness lives on in #330.

[ligon]

Resolved by #447 (merged to development). 4/5 fixed; the 5th (api_vs_replication[food_quantities]) is the pre-existing parallel-cold flake tracked in #330, not a stale baseline. Post-fix full cold gate: 0 failed (was 5).