You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
chore(deps) — bump next from 16.1.6 to 16.4.1 (#563)
chore(ci) — dependabot.yml and CODEOWNERS for the security surfaces (#441) (#567)
docs(ops) — the twelve non-main branches, triaged by what proves each removal safe (#560)
docs(agents) — ## Numbers points at docs/GATE.md instead of copying the figures (#559)
docs(gate) — re-measure the baseline over 2026-11-04 — 156 s a run, 176 s to the last job (#557)
Known limits
Account deletion needs a live session. The control lives on /settings and acts on the signed-in person. Anyone whose managed-auth user is already gone cannot self-serve it, and there is no operator tool in the product for that case.
A failed scrub is resumable, not automatic. When the auth deletion succeeds and the sweep then throws, the intent survives in account_deletion_job, an operator finishes it with the command in docs/ACCOUNT-DELETION-SPEC.md, and scripts/db-verify.mjs fails on a row still open after 24 hours. That is a check somebody runs, not a pager: nothing alerts on its own and nothing retries automatically, so an abandoned row is visible only to whoever runs the verify.
The marketing opt-out row keeps the email address on purpose. It is the one retained identifier, it exists so the address is never mailed again, and the privacy policy says so rather than claiming no address is kept. Billing records (7 years) and security audit entries (24 months) carry no name or address.
A shared assessment survives the deletion of one participant. The other person keeps their report; the deleted person's seat reads "Deleted account" and their answers remain inside that completed report with the identifying prose rewritten. That is the product decision, recorded in docs/ACCOUNT-DELETION-SPEC.md.
The dependency cohort moved major versions with no application change. react 19.3.1, recharts 3.10.1 and swiper 12.2.0 all landed without source edits; the carousel screenshots in docs/qa-swiper-12/ are the visual evidence, and the two pairs that differ are autoplay timing rather than regressions.