[vahid-ahmadi]

Cause

The scheduled Fetch data vintages run 42980628369 (main e84d8c19) stopped at the pages step:

$ python3 economy/topics.py
  File "economy/topics.py", line 2290, in us_inflation_note
    change = 100 * (now["value"] / year_ago["value"] - 1)
TypeError: 'NoneType' object is not subscriptable
FAILED (1): python3 economy/topics.py

The same run's fetch step shows written us_cpi: FRED published the 2026-10 CPIAUCSL index (337.476). year_earlier() looks for 2025-10, and the store has no such row. That is not a fetch bug: FRED itself publishes no value for that month. The row is blank in the source CSV:

$ curl -s 'https://fred.stlouisfed.org/graph/fredgraph.csv?id=CPIAUCSL&cosd=2025-08-01'
2025-09-01,324.245
2025-10-01,
2025-11-01,325.063

fetch.py drops the blank, as intended. The recovered vintages (#234) are not involved. This is the first month whose year-earlier counterpart is that hole, so it surfaced only now. It recurs only for 2026-10: from 2026-11 every month has a counterpart again.

Three places were affected, not one:

• us_inflation_note crashed (the failure above).
• stat_card(us_cpi, "yoy") would have raised RuntimeError next.
• us_inflation_hook printed the 2026-09 rate beside the 2026-10 index, because yoy() skips 2026-10 but the hook took latest().

Fix

• New latest_year_on_year(series) returns the newest observation that has a counterpart a year earlier, that counterpart, and any newer observations that have none.
• The US inflation note, hook and card, and the US growth note, use it. Where the newest month has no counterpart, the page says so: "The newest stored month, 2026-10, has no year-on-year rate here: the store holds no 2025-10 observation to compare it with, and none is interpolated." No number is derived for 2026-10 from any other month.
• movement(): from 2026-11 the "prior month" rate would really be 2026-09's. The text now names that comparison, against 2026-09, the latest earlier rate (2026-10 has none), instead of calling it "the prior month".
• When every month has a counterpart, the output is byte-identical: python3 economy/topics.py --check passes against main's committed pages.

Tests

tests/test_topics_year_on_year.py uses a synthetic series with the same hole, because main's data ends at 2026-09 and a test reading the store would pass before the fix.

• On main (before): 7 of 8 fail. The note test fails with the production TypeError, the card test with RuntimeError, and the hook test because the hook pairs the wrong index.
• On this branch: 8 of 8 pass.

Pipeline check. The full pipeline ran locally on today's real data: data/fetch.py at about 10:03 UTC wrote us_cpi 2026-10, then data/refresh.py --carry-from data/vintage-refresh --skip-fetch ran with this fix.

• Every step and every gate passed: pytest tests gave 3219 passed and 87 skipped, and the append-only check passed.
• None of that fetched data is committed here. This PR changes only economy/topics.py and the new test file.

Once this merges, dispatching Fetch data vintages on main rebuilds #195.

🤖 Generated with Claude Code