A document_attachments row is reachable from eight places. A row referenced by
none of them is stored, retained for seven years under BFL, and connected to no
bookkeeping at all. Nothing surfaced those, so they accumulated: 4 497 across
210 companies, 481 of them in the preceding week.
The naive predicate is a trap. Without a mime filter the same query returns
15 806 rows, and 11 309 of those are archived PSD2 bank-API responses that are
unlinked by design. Putting them on an orientation surface would hand an agent
eleven thousand items of work it must not do, which is worse than showing
nothing.
So the rule is an allow-list of the mime types an underlag can actually be.
Measured on production: application/json was 11 309 of 11 309 PSD2 archive, and
pdf/png/jpeg/heic were 0 of 4 495. The split is clean, and an allow-list keeps
the next machine-payload format out by default rather than after someone
notices it leaking.
Two passes, mirroring fetchPurchasesWithoutUnderlag: the indexed column filter
first, then eight reference lookups that run only when candidates exist, so a
company with none costs exactly one query. The scan cap is set by URL length
rather than table size, because every candidate id is echoed back through those
eight .in() lookups.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Jakob Wennberg
5fe3ae71a1