* fix(flutter): find snapshot headers that a read chunk boundary splits
Within one chunk the snapshot hash scanner resumes four bytes after each
magic, but across chunks it resumed after the whole 52-byte record and
dropped the carried bytes. A magic starting 49-51 bytes after a rejected
(non-hex) record was therefore lost when its four bytes straddled a read
boundary, so the result depended on chunking and a conflicting hash could
be omitted while coverage stayed complete. Resume after the magic in both
cases.
* refactor(flutter): drop the snapshot scanner cursor the fix superseded
After a complete record the carried tail is shorter than a magic, and a
pending magic has not been processed, so every push can rescan its carried
bytes from the start without counting a record twice. Remove the absolute
cursor and base offset, and cover sections 49-51 bytes after a rejected
record with two-chunk and byte-at-a-time feeds.