Every file the walk records now gets three hashes in the hash phase:
head and tail, of its first and last 64 KiB, and content. The content
hash covers the whole file under 100 MiB, and 1 MiB at each gigabyte
from 100 MiB up, so a large file is never read in full. One 100 MiB
constant replaces the 10 MiB and 50 MiB thresholds. The size census,
the records stored without hashes and the content phase are gone, so a
scan has four phases and a rescan reads only new and changed files.
report and trees no longer special-case a missing content hash. README
"Duplicate detection" and "Rationale" are rewritten.
Judgement call: info drops its two hash-count lines, which would now
always equal the record count.
Model: opus-5-5
scan never records or reads anything on a filesystem that is not
persistent (tmpfs, proc, sysfs and the like), however it reaches it; no
flag turns this off. It notes each such filesystem's device ID as it
starts: on Linux from /proc/self/mountinfo, on macOS by lstat of each
listed mount point. Below an operand the walk also skips the default
excludes, a path list per system (Linux, macOS, and Windows, not used
yet); --no-default-excludes turns off only those. A default exclude that
is itself an operand gets no skip line. The tests stop at once, saying
why, when TMPDIR is not on a persistent filesystem.
Judgement call: an unreadable mount list fails the scan, and no flag gets round it.
Judgement call: a file on such a filesystem below an operand gets the same line as a directory.
Judgement call: such an operand that another operand's walk also reaches is named twice.
Unverified: whether macOS's mount list gives the device lstat reports, so macOS keeps lstat, and a hidden mount there can still add another filesystem's device.
Unverified: the macOS code is never built or run here; bind mounts need root to test.
Model: opus-5-5
clawbot