[yannickmonney]

Summary

The database entrypoint only ever created the ready marker /tmp/.db_ready; nothing cleared it. /tmp outlives a stop whenever the runtime keeps the container's filesystem, as docker restart does, so a restarted database read the previous start's marker and reported ready (HEALTHCHECK, compose waits, CLI and CI waits) as soon as Postgres answered, before this start's init scripts and migrations had run.

The fix clears the marker with a top-level rm -f /tmp/.db_ready before the background init block, so the marker only exists once this start's init has finished. services/db gains a test script and a Bun test for the entrypoint.

Checklist

• bun run check passes — the new and touched parts pass; on the test VM the remaining failures (@tale/cli#test, @tale/platform#test:ui, @tale/session-manager#test) are the same on a clean origin/main worktree there (VM environment: no docker compose plugin, memory, timing) and none depends on @tale/db. CI is the reference.
• bun run lint:sast — N/A: a shell line and a test; no SAST-relevant surface.
• Translations — N/A: no user-visible string.
• Docs — N/A: the documented behaviour ("Wait for initialization": wait for /tmp/.db_ready) stays true and now also holds after a restart.
• README — N/A: no change in usage.

Test plan

• bun test services/db: 2 tests pass (run 3 times). A structural test checks that the clear is a top-level line before the background block and its touch; a behavioural test runs the real entrypoint twice under bash with a fake server: on the restart the server starts with no marker, and the marker returns once init has finished. With the fix reverted, both tests fail (the restart reads the stale marker).
• shellcheck, oxfmt, oxlint and a strict ad hoc tsc on the new test: clean. knip, format:check, lint:conflicts, lint:links: clean. commitlint: clean.
• Real container: built the db image from main and from this branch, ran each, then docker restart. On main the marker was present in every sample taken while pg_isready answered and before the second Database ready.; with the fix it was absent in every such sample and present after init finished.