e2e: Mailosaur deleteAllMail() is server-wide and can delete a concurrent run's verification email

Found during the adversarial review of the e2e suite; not addressed in #1572.

`tests/mail.ts` → `deleteAllMail()` calls `client.messages.deleteAll(serverId)` from `afterAll`, which deletes **every** message on the Mailosaur server, not just the ones this run generated.

That was safe when one lane ran at a time. It no longer is:

- the trusted lane (daily + 15-min stage poll) and the PR lane can overlap — `host/run-sandbox.sh` explicitly notes both drive it;
- a local `pnpm test` on a developer machine uses the same Mailosaur server.

If one run's `afterAll` fires while another is inside `waitForVerificationEmail()` (a 90s window), the second loses its message and fails with "No verification link found" — an unreproducible red that triage will most likely classify as a regression.

Fix ideas:
- Delete only messages addressed to the address this run generated (`client.messages.list({ sentTo })` → per-id delete). Addresses are already unique per run, so that is sufficient.
- Or give each lane its own Mailosaur server and keep the server-wide purge.

Start in tests/mail.ts with deleteAllMail() and trace how sentTo is generated and used by waitForVerificationEmail(). Review host/run-sandbox.sh for the overlapping trusted and PR lanes. Done means cleanup removes only this run's messages, while another concurrent run can still retrieve its verification email; add or run the relevant e2e coverage if the existing suite exposes it.

Written by the indexing model from the issue text.