fix: drain legacy logs fully on the first post-migration distill

[scode]

Found by the live E2E battery, not by review: the migration flow promises the first post-migration leiter distill
drains the legacy logs and removes the directory, but it actually took two runs — scan-time obsolete cleanup only
deletes files below last_distilled as of scan start, so the log emitted by the run itself always survived to the
next cycle. The remote suite's migration step failed on a real box exactly there.

The fix is a post-commit sweep: after last_distilled has provably advanced, delete legacy logs strictly below the
new cutoff, then remove the directory once empty. Post-commit placement keeps the crash posture identical to
scan-time cleanup (a failed run's sweep never executes), and the strict cutoff mirrors the inclusive >= scan rule so
nothing is deleted before a run that emitted it has committed. Review of the fix surfaced one real interaction: a
concurrent distill's sweep can now delete a legacy log another in-flight run had listed but not yet read — that read
becomes warn-and-skip rather than a spurious hard failure, since the vanished file's content is committed by whichever
run swept it.

changelog: include