[hwanjjang]

Summary

full_response::a_province_under_its_new_name_finds_its_town_and_mid_forecast failed intermittently. It looked time-of-day dependent, but the cause is a race in the test:

• S3Source::zones() starts the process's forecast-zone fetch without waiting for it (src/domestic/kma/source/mod.rs ~462-514). refresh_zones() then uses the stored list if there is one, otherwise the release's land zones.
• The test's double answered the zone page with Seoul only. When that list was stored first, towns outside Seoul had no city zone, their mid temperature was not collected, and the test failed with temperature-unavailable.
• Forcing the wall clock changed nothing: the backend and the zone path use the injected time.

Failure rate, measured by running the test repeatedly with a faked wall clock:

Code	Rate
master	1/20 at 08:15, 3/20 at 20:15
#2872 head	about 60%

#2872 reads day products at the start of a request, so the zone read starts later.

Change (test only)

• The double answers the Seoul fixture plus every land zone of Zones::packaged(), as the provider's list does.
• The test runs each town both with the list stored first and without it.
• Red: 5/5 failures at forced 08:15 and 20:15 before the fix. After: 20/20 in a row, 5/5 with the CPU saturated, and 60/60 across six forced times.

The same change is in #2872, declared under that PR's task.

Checks

• fmt and clippy are clean. The full suite has 1,213 passed and 0 failed.
• The placement gate (KMA-TEST-CLOCK.json) and the artifact policy pass.
• Other tests that read the real clock are only at risk across a boundary (KST midnight or an hour), not at a time of day: tests/domestic/quota.rs:50, source_s3.rs:1083, air_s3.rs:458 and :628. They are listed, not changed.

🤖 Generated with Claude Code

[hwanjjang]

Cross-provider review (Codex gpt-6.1-sol, high), round 1: PASS at 0c76e77.

Recommended

• TC-force-fallback-order: with stored_first = false, the packaged fallback is not guaranteed to run first, because the zone fetch can finish first. Selected: the zone request will be held at a barrier in that branch. The identical test change goes into server2: restore station observations from day products after a restart #2872, which carries the same test.

[hwanjjang]

Cross-provider review (Codex gpt-6.1-sol, high), round 2: PASS at 64541ef. TC-force-fallback-order is resolved: the held zone request now pins the packaged fallback path.

[hwanjjang]

Final readiness (pre-merge): merge-ready at 64541ef on master 0ed2eaa. CI passes, and the cross-provider review passed (Codex, two rounds; the selected recommendation is applied). The PR has no auto-merge and no merge queue. Merging is AK's decision. If master moves before this is merged, master must be merged into this branch before any further push (the CI placement base stays at the merge base). #2872 carries the identical test change.