# Fedora / RHEL
sudo dnf install gcc make libatomic pkgconf-pkg-config \
    glib2-devel sqlite-devel xxhash-devel \
    libuuid-devel libmount-devel libblkid-devel libbsd-devel

# Debian / Ubuntu
sudo apt install build-essential pkg-config \
    libglib2.0-dev libsqlite3-dev libxxhash-dev \
    uuid-dev libmount-dev libblkid-dev libbsd-dev

Requires Linux 3.13+ with a btrfs or XFS filesystem. On XFS, filesystem identification is unprivileged on Linux 6.4+; on older kernels run oans as root (a scheduled root job is unaffected).

• Hardlink hashfile-emptying hazard fixed. INSERT OR REPLACE on UNIQUE(ino, subvol) could cascade-delete rows for other links to an inode and silently empty the hashfile while exiting 0; an in-memory seen_inodes guard and a regression test pin it.
• Use-after-free & leak fixed when a rejected hashfile is recreated (a closed handle was handed back to the caller); found by running the whole suite under valgrind, which is now a CI job.
• Dedupe robustness: stop spinning when a round makes no progress (upstream #396/#407), clamp the source range to file size to avoid EINVAL, fix the infinite loop on NoCOW files (#376), and skip members whose size changed since the scan.
• Sparse-file fixes: handle files with a trailing hole (#374); hash the actual block, not a stale zero check.
• Uninitialised-memory fix in UUID config load, with a committed valgrind suppressions file for the one library false-positive.
• In-memory (no --hashfile) “database table is locked” fixed (shared-cache read/write connection split).
• Robust output: 32-bit-correct counters, sanitized filenames in the status line, no -nan% bar.
• busy_timeout so transient lock contention retries instead of failing.
• Report modes open the hashfile read-only — --stats/--history/--json are safe to run while a scan or dedupe is in progress.
• Refuse unsupported-fs roots up front instead of silently storing 0 files; hint at permissions when a hashfile can't be opened; report a version even when built outside a git checkout (#387).

• Skip already-shared files up front (#331) — the headline warm-rescan win.
• No cross-generation reprocessing of duplicate groups that span passes (halves kernel dedupe traffic, fixes accounting).
• Fast dedupe on trees larger than RAM: keep just-hashed data in the page cache for the dedupe phase and prefetch each FIDEDUPERANGE round, so the kernel byte-compares from RAM instead of a slow cold re-read — ~13× faster than upstream under memory pressure (benchmark).
• Batched SQLite transactions (~10 s cadence) for both the change-detection reads and writes — hundreds of thousands of per-file lock syscalls collapse to a few hundred (~24 % faster rescans).
• Parallel directory walk (--io-threads), capped where btrfs metadata contention plateaus (~8 walkers).
• Compact 64-bit path-hash index instead of the full-path index.
• Largest-files-first (LPT) hashing to shorten the idle tail.
• fiemap streamlining: ranged queries in the dedupe phase, one shared core, a single ioctl for the rescan extent, and a resume cursor that kills an O(n²) rescan.
• Skip all-hole blocks in sparse files; grow the scan's hash arrays geometrically, not one element at a time.
• Fewer redundant lookups: cache per-device fs verification and btrfs subvolume lookups; build find-dupes indexes after the scan; drop indexes that duplicate a UNIQUE prefix.
• Lower peak RSS: SQLite cache 256 MB → 64 MB per connection, per-role cache budgets, smaller read buffers, and an open-addressing seen_inodes set instead of a GHashTable.

• --stats — a full hashfile report: file/hash counts, logical duplication ratio, file-size summary, free space a VACUUM would reclaim, and the top duplicate groups (size × copies, with an example path).
• --history — a per-run timeline of space actually reclaimed, plus lifetime totals, from an appended run_history table.
• --json — a flat metrics object for jq / Telegraf / dashboards.
• --progress=json — one JSON object per phase (~1/s) to stderr, ending with a done event, for monitoring scheduled runs; stdout untouched.
• Scan diagnostics (DUPEREMOVE_SCAN_STATS) — queue starvation and write-lock contention counters.

• Unified live progress UI across scanning / hashing / dedupe / done — one stable, colorful status line with rate, percentage, runtime and a weighted-ETA, a mapping: pre-read phase, and a summary block.
• Honest reporting: the summary reports disk space actually freed (one physical copy kept per group), not an inflated shared-extents figure.
• Progress polish: ellipsize the path middle so the numbers stay visible, always-human sizes, no bar pinned at 100 %, no live-block drift, files-examined shown during listing, quieter per-file errors, and “Skipping dedupe” no longer collides with the status bar.
• Self-contained --help/usage instead of shelling out to man.

• Self-describing hashfile: each run stores its options, roots and excludes, so a bare oans --hashfile=FILE replays the last run incrementally — the basis for zero-config scheduling.
• systemd oans@ service/timer templates (make install-systemd) for weekly, idle-priority dedupe, with a NAS quick-start guide.
• --autotune empirically picks the fastest --io-threads for the backing storage (and a device-type heuristic sizes the default for HDD/RAID pools).
• --min-filesize, --no-color and -q added; the legacy --fdupes mode and other dead/testing options removed.
• Automatic housekeeping: prune deleted files from the hashfile after a scan (stat-based), then VACUUM when a build or prune left it worth it.
• Branded hashfile (SQLite application_id, format 5.0) so oans and duperemove never misread each other's caches.
• Drop-in compatibility: make install adds a duperemove symlink and keeps the stable net change in shared extents output line.
• Packaging & release: prebuilt x86_64 tarball attached to each GitHub release (scripts/release.sh), AUR PKGBUILDs, and a hand-drawn logo + scripted demo GIF.
• Build & CI hardening: src/ layout, -Wextra/hardened flags, reproducible builds, and CI running the unit + integration suites (and a valgrind pass) against real btrfs and XFS.