pack:verify: await the --web child's exit before removing its work dir #1826

Follow-up to #1801 (fixed in #1814). Same shape, lower risk — filed so it isn't rediscovered from scratch if it ever does bite.

Context

#1801 was smoke:tui failing intermittently with ENOTEMPTY: done() sent SIGTERM and then synchronously rmSync'd the work dir, without waiting for the child to exit. scripts/pack-and-verify.mjs has the same kill-then-remove pattern in two places.

The two sites

scripts/pack-and-verify.mjs:330 (success path) — the closer analogue. verifyWeb()'s finally { stop() } sends SIGTERM (line 377-380), control returns to the main flow, and rmSync(work, …) runs immediately with nothing awaiting the child's exit.

scripts/pack-and-verify.mjs:74-78 (fail()) — kills webChild and rmSyncs workDir back to back.

Why this is low priority, not a live flake

• Unlike smoke-tui, the --web child's HOME is not redirected into the work dir — work is only its cwd and the location of the installed package it reads. In smoke:tui intermittently fails with ENOTEMPTY cleaning its temp dir #1801 the work dir was HOME, so the Ink app was actively writing config/storage into the directory being deleted, which is what made the race fire. Nothing here writes into work on shutdown, and on macOS a live cwd doesn't block removal.
• fail() runs only when the verify has already failed and is exiting 1, so an ENOTEMPTY there cannot manufacture a spurious red — worst case it masks the real failure message with an rmSync stack and skips the "tarball retained at …" hint.
• pack:verify is not in the local pre-push gate. It runs in CI on the publish path only (.github/workflows/main.yml:155), so a flake would be rare and loud rather than the learned-to-re-run red smoke:tui intermittently fails with ENOTEMPTY cleaning its temp dir #1801 became.

Suggested fix

Apply the #1814 treatment to stop(): make it await the child's exit event (escalating SIGTERM → SIGKILL after a grace period, then proceeding anyway with a warning) before the caller removes work, and wrap the rmSync calls so a leftover temp dir warns instead of throwing over the real failure message. See scripts/smoke-tui.mjs after #1814 for the pattern.