Published
- Month arithmetic: Go overflows, MySQL clamps — same intent, five days apart
- The LEFT JOIN column that only fails at runtime — Error 1054, and the silent NULL next to it
- The 1fr that refuses to shrink — a 1509px track in a 900px box, and an 880px void from a missing wrapper
- The certificate that renewed and still expired — disk date vs served date
- RowsAffected, and the upsert that explodes — 0 changed rows read as "row missing"
- The unique index a deleted row still holds — projecting dead rows to NULL
What counts as publishable here
Three conditions, all of them checked before a page goes up. It has to fail on a machine we can run, so the evidence is watched failing rather than remembered. It has to leave a trace we can paste - the compiler error, the row count, the certificate the server actually served. And it has to end with a check that would fail if the fix were reverted, because "it works now" is not a criterion. When the snippet is useful outside our own systems, it becomes a runnable reproduction in a public repository.
Runnable reproductions
- mysql-rowsaffected-behavior - a two-table script showing an upsert that changes nothing while matching a row.
- soft-delete-aware-unique-constraint - the generated-column trick that lets a deleted row stop holding a unique key.
- verification-gate-integrity - a gate script that exits 0, 1 and 2, so a green run means something.
Every page here ends with how to know it worked and how to undo it. If we cannot reproduce the failure on this machine, it does not get published - and a page that only cites documentation says so at the top.