A partial build output at the target destroys or strands the good .prev DB

Found reviewing #633 (round 7). Data loss, verified both branches. `_clear_build_target`'s orphan recovery only fires when the target is absent or a symlink. A killed `semsql make` leaves a **partial real file** at the target, which takes neither path: ``` partial >= min_size -> DbEnsureResult(usable=True, built=False) | builds=0 | good .prev survived: True partial < min_size -> DbEnsureResult(usable=False, built=False) | builds=1 | good .prev survived: False ``` - **partial ≥ min_size**: the partial is silently accepted as the DB (no build runs at all) while the good DB stays stranded at `.prev`. - **partial < min_size**: `os.remove(kept)` deletes the good `.prev` before `os.replace(db_path, kept)` moves the partial into its place. The build then fails and the code prints *"Restored the previous ncbitaxon.db after the failed build"* — having restored the partial. The good 13 GB DB is gone, with a message claiming the opposite. `_clear_build_target` should never overwrite a usable `.prev` with an unusable target, and should treat "partial file at the target + usable .prev" as the recovery case it already handles for "absent target + .prev". Related: the `except BaseException` comment claims SIGTERM protection. Python installs no SIGTERM handler, so the process dies without unwinding and the `finally`/`except` never runs — this is the ordinary interrupted-build path, not an exotic one, and the comment should not imply otherwise. 🤖 Generated with [Claude Code](https://claude.com/claude-code)