fix(migration): backfill workspace.time_used with current timestamp

[kitlangton]

Summary

• Existing workspace rows currently get time_used = 0 after the v1.14.43 migration shipped in fix(storage): default workspace time migration #26556, so they sort as the oldest "most recently used" rows forever even when they were in active use before the upgrade.
• Schema declares time_used: integer().notNull().$default(() => Date.now()), so newly inserted rows via Drizzle get a real timestamp — only the migrated rows are wrong.
• Ship a new migration 20260509205313_backfill_workspace_time_used that runs UPDATE workspace SET time_used = unixepoch() * 1000 WHERE time_used = 0.

Why a new migration (and not edit the existing one)

Drizzle's bun-sqlite migrator filters getMigrationsToRun by name only, not hash:

function getMigrationsToRun({ localMigrations, dbMigrations }) {
  const dbNamesSet = new Set(dbMigrations.map((m) => m.name).filter((n) => n !== null))
  return localMigrations.filter((lm) => !lm.name || !dbNamesSet.has(lm.name))
}

So editing 20260507164347_add_workspace_time/migration.sql in place is a no-op for any user who already applied the prior DEFAULT 0 version (i.e. anyone on v1.14.43). A fresh migration with a new timestamp is the only path that reaches them.

Why WHERE time_used = 0

• Users on v1.14.41 or v1.14.42 (broken migration crashed): the ALTER migration runs first with DEFAULT 0, every row is 0, the backfill catches them all.
• Users on v1.14.43 (DEFAULT 0 fix applied): only the new migration runs, every row is still 0 from the prior backfill, the backfill catches them all.
• Future / dev: the migration is a no-op for any row that already has a non-zero time_used. Idempotent and self-documenting.

Why unixepoch() * 1000 (not unixepoch('subsec') * 1000)

• Whole-second precision is sufficient for "most recently used" sort ordering.
• Avoids the float-cast truncation race that unixepoch('subsec') * 1000 introduces vs JS Date.now().
• Matches the millisecond-integer convention the rest of the codebase uses.

Verification

• bun test --timeout 5000 test/storage/workspace-time-migration.test.ts — 2 tests, covers both fresh-upgrade and already-DEFAULT-0 paths.
• bun typecheck

[kitlangton]

Closing — on reflection, DEFAULT 0 (shipped in #26556) is the better behavior here.

time_used has exactly one reader: the recent-workspaces sort in dialog-workspace-create.tsx. Backfilling pre-existing rows to unixepoch() * 1000 would cluster every migrated workspace at the top of the recent list with identical timestamps, drowning out real recent activity. Leaving them at 0 lets them surface naturally the next time someone opens a session in them (which bumps time_used to Date.now() via projectors.ts).

The original 'default to now instead of nullable' framing was based on a misread of the shipped fix — the column is already NOT NULL DEFAULT 0, not nullable.

No-op.