Track objects are named <prefix>_<trip_id>.parquet. The id is derived from
the configured prefix rather than matched with a generic pattern, so a name
in an unexpected layout yields the whole name — visibly not a trip id —
instead of being silently returned as one.
That distinction is what the error below is for. The previous
implementation, gsub(".*_([0-9]+)\\.parquet$", "\\1", filenames), returns a
non-matching name unchanged, so a bucket written under any other naming
convention reads as "none of these trips are stored" and
ingest_pds_tracks() re-fetches the entire history from the PDS API. Timor
hit exactly this: 0 of 99,219 stored tracks recognised, and a first run that
would have re-downloaded 98,472 of them. The failure was silent, expensive,
and indistinguishable from a fresh bucket.