export ts --stitch --linger exits on internal error when its catalog request lands on a publisher that just died #5347

A concrete case for quest/m1/dropped-sources.md,
with one consequence the quest doesn't list: ts::Follower exempts Dropped from failing loud, but a
drop forwarded by a relay arrives as internal error, so the exemption can only match locally.

On main at 90500867d, moq export ts --delay 1s --linger 60s --stitch on the default
moq-lite-06 exited 1 with moq: internal error in 1 of 4 crash-replace runs; the other three
resumed at every replacement. Rig: one relay on loopback, the export started first, then three
moq import ts sessions of one broadcast, 15 s each, the first two SIGKILLed 5 s before their end and
each replaced 3 s later. Default idle timeout.

• Session 0 is killed at 10:22:34.0; session 1 connects at 10:22:37.3 and is killed at 10:22:47.20.
• At 10:22:46.95 the relay times session 0 out and ends the export's subscriptions; the export logs
  broadcast ended, waiting for it to return at 10:22:47.26 and sends a catalog track info request.
• The relay routes it to session 1, dead for 60 ms but not yet timed out. At 10:22:57.26 it times
  session 1 out and logs track info error … track=catalog.json err=dropped; the export logs
  track info failed … err=internal error and exits. Session 2, up since 10:22:50.5, would have
  served the return.

Error::Dropped has no arm in From<&Error> for StreamError, so it goes out as Internal, and
State::Following in rs/moq-mux/src/container/ts/follower.rs treats any resolve error but
Unroutable | Dropped as a refusal from a live instance. If the quest ends this path with
Unroutable, as it plans for requests a front aborts, the follower resumes here unchanged; the
quest's regression could cover a request in flight when the upstream session times out. On
moq-lite-07 the replacement's epoch ends the old broadcast first, so we did not see it there.

(Written by Claude Opus 5.5)