fix: --check measures lag on write time, not on data age

[mspinola]

--check flagged six entries on your real store, none of which were a problem. All six were old-but-correct, and the check had no way to tell that apart from a failure.

The wrong question

The check exists to catch a partial run: one entry failing while its peers succeed. It compared each entry's last_date against the domain's newest last_date, which measures how recent the data is. That is a different question.

Two legitimate classes of old-but-correct entry were flagged permanently:

Thin markets that drop out of the CFTC report. Below the reporting threshold a contract stops appearing and returns months later. Measured over 19 years of weekly data:

NKD_240741: 19 gaps >10d, longest 168d  (recent: 168d, 147d)   currently 140d
ZO_004603:  22 gaps >10d, longest 294d  (recent: 294d, 105d)   currently  49d

Both are comfortably inside their own normal.

Retired hist_code contracts, frozen by design. These were suppressed by name via hist_code_names(), except the store also holds bare-code duplicates (23977A alongside RTY_23977A) written before the naming convention gained a symbol prefix. The suppression set only ever generated the prefixed form, so the bare ones warned forever.

The right question

Both classes are rewritten on every producer pass, so both carry a current updated_at:

NKD_240741  last_date 2026-03-03  updated_at 2026-07-26T16:13:07Z
ES_13874A   last_date 2026-07-21  updated_at 2026-07-26T16:13:07Z   <- identical

Comparing updated_at against the domain's newest write separates "the source had nothing newer" from "the producer skipped this". It also makes the bare-duplicate naming mismatch moot, rather than needing the suppression set to cover two conventions.

A hole found while testing

When no entry had a parseable updated_at, the new shape initially reported everything as current. Turning "cannot check" into "all clear" is the precise failure this check exists to prevent, so unusable timestamps are now flagged. There is a test for it.

Verification

On the real store: 6 false positives to 0.

✓ every entry was written by the latest producer pass.

And a simulated skip is still caught:

⚠ cot_legacy: 1 entry NOT WRITTEN in the last run (>3d behind the domain's newest write):
    ES_13874A         last written 2026-07-01T04:00:00Z  (25d behind)

Suite 129 passed, ruff clean.

Note

hist_code_names() and the ignore_lag parameter are kept and still work. They are no longer load-bearing for this case, but they remain the escape hatch for suppressing an entry by name.