A previously aborted ec.rebuild leaves 0-byte placeholder files for the
shards it was about to write. On the next attempt, findShardFile saw
those ghosts as "present", the read loop returned nil on the first n==0,
and rebuild claimed success while leaving the volume mounted with empty
stripes — manifesting downstream as "too few shards given" or silent
corruption.
Now: skip 0-byte files when scanning, validate that all surviving shards
agree on size, bound the read loop by that size instead of bailing on
the first short read, and remove residue after a successful rebuild so
it can't shadow real shards next time.
Chris Lu
b122308bf4