sfdupes/TODO.md

# Workflow

- Take an issue from the `1.0.0` milestone on the tracker. Work that is not on
  the tracker yet is filed as an issue first.
- Branch from `next`.
- Do the work, with tests, in small focused commits.
- Record the step at the top of Completed Steps, in the same commit as the work.
- Push the branch and open a PR against `next`. Its title ends with
  ` (closes #N)`.
- An independent review gates each merge to `next`. Every finding is addressed
  or rebutted on the PR.
- Only the owner merges `next` into `main`.

# Status

- Pre-1.0.
- The tracker is authoritative for the pre-1.0 backlog. The open issues under
  the `1.0.0` milestone are what remains before the tag. This file records
  history and process, not the queue.

# Next Step

- Take the next issue from the `1.0.0` milestone,
  https://git.eeqj.de/sneak/sfdupes/milestone/17. The issues are not restated
  here, because a copy drifts as soon as the tracker moves.

# Completed Steps

Newest first. An entry ends with its date and its issue or branch, where those
are known.

- The three disk size tests in `scan_test.go` no longer fail at random on ZFS.
  `unix.Sync` can return before ZFS has counted a file's blocks, and one ZFS
  commit can take a file through several sizes. The test that needs the size to
  change waits for it, for at most 30 seconds. Each test reads the size before
  and after its scan. The rewrite test rescans until the two reads match, then
  compares exactly. The other two fail only on a `disk_size` outside the two
  reads (2026-11-02, https://git.eeqj.de/sneak/sfdupes/issues/150)

- `TODO.md` rewritten from scratch by the same rules as `README.md`, sentence by
  sentence. Every completed step is kept, with its date and its issue or branch.
  The policy-compliance checklist of 2026-07-22 is folded into its step
  (2026-11-02, https://git.eeqj.de/sneak/sfdupes/issues/135)

- `README.md` rewritten from scratch, sentence by sentence, as a newcomer reads
  it. It leads with what the program is and how to use it. A reference section
  covers every setting and every subcommand. Development and testing come last.
  Every statement was checked against the code on `next`. Five comments in
  `main_test.go` cite "Settings", where their rules now live. One comment each
  in `scan_test.go`, `trees_test.go` and `exclude_test.go` cites what the README
  now says (2026-11-02, https://git.eeqj.de/sneak/sfdupes/issues/135)

- Each record holds `disk_size`, the bytes the file takes on disk. That is its
  allocated blocks, `st_blocks` × 512, from the `lstat` call the walk already
  makes. `trees` prints a `disk_size` column after `size`. It is the sum over
  the `dupe` copy's files. On a compressed or sparse file `disk_size` differs
  from `size`. A rescan updates the `disk_size` of an unchanged file whose size
  on disk changed, without reading the file (2026-10-27,
  https://git.eeqj.de/sneak/sfdupes/issues/147)

- The database keeps the number of records and the sum of their sizes in the one
  row of a `totals` table. Three triggers on `files` keep that row current.
  `info` reads that row. So do `scan`'s `load` phase and the `read` phases of
  `report` and `trees`, for their totals. No count reads every record any more.
  The triggers make each write to `files` slower. `scan` keeps SQLite's
  temporary data in memory. That wins back part of that time (2026-10-27,
  https://git.eeqj.de/sneak/sfdupes/issues/146)

- an operand at or inside a directory named `.zfs` is never scanned, whether
  given alone or under another operand. `scan` warns about it, counts it as
  skipped and does not walk it (2026-10-26,
  https://git.eeqj.de/sneak/sfdupes/issues/142)

- A file or directory removed while the scan runs is not an error, in the walk
  or in the hash phase. The scan prints no line for it and does not count it as
  skipped. When the first of several hard links to a file is gone, the hash
  phase reads the next one instead. On a terminal every progress display starts
  with the spinner, the hash phase's included. A phase with an exact total ends
  its line with its percent and a 10-column bar at the terminal's right edge.
  The bar no longer moves as the line's length changes. A line too wide for the
  terminal now loses its rate before its bar. The time remaining is labelled
  `remain`, not `eta` (2026-10-26, https://git.eeqj.de/sneak/sfdupes/issues/133)

- The `head`, `tail` and `content` columns are BLOBs, each holding the raw 32
  bytes of a SHA-256. That is half the size of the hex text they held. The
  hashes stay raw from the hash phase through the database to `report` and
  `trees`. Reading a record whose hash is not 32 bytes is an error (2026-10-26,
  https://git.eeqj.de/sneak/sfdupes/issues/134)

- The hash phase gives three hashes to every file that the walk records. `head`
  and `tail` are the SHA-256 of its first and last 64 KiB. `content` is the
  SHA-256 of the whole file under 100 MiB, and of 1 MiB at each gigabyte from
  100 MiB up. Gone are:
    - the 10 MiB and 50 MiB thresholds;
    - the size census, which counted the sizes of every walked file and of every
      record outside the operands. A file was hashed only when the census showed
      another file of its size;
    - the records without hashes;
    - the content phase.

    A scan has four phases, and a rescan reads only new and changed files.
    `info` no longer counts the records with `head` and `tail` hashes, nor those
    with a content hash, since every record has all three (2026-10-25,
    https://git.eeqj.de/sneak/sfdupes/issues/132)

- `REPO_POLICIES.md` and the four scripts that pass `CACHEBUST` were copied
  again from `sneak/prompts` `next` at `0653579`. Two steps now come before
  `ARG CACHEBUST`:
    - the module download in the `lint`, `test` and `builder` stages of the
      `Dockerfile`;
    - the package install in `builder`.

    Docker keeps them until one of these changes:
    - the base image;
    - the install line;
    - `go.mod` or `go.sum`.

    Every step that checks or compiles the code still comes after
    `ARG CACHEBUST`. `golangci-lint run` also keeps the linter's own cache.
    README says the same (2026-10-18,
    https://git.eeqj.de/sneak/sfdupes/issues/109)

- The release rule was set. Until 1.0, each tip of `main` that passes the build
  gets the next `v0.0.N` tag. `go install sneak.berlin/go/sfdupes@latest` then
  installs it. `go.mod` retracts `v0.0.1`, the early build without a database.
  When the linker set no version, `sfdupes --version` prints the version Go
  recorded in the binary. That is the tag, or the pseudo-version of a
  `go install` build. It prints `dev` only when neither is known. README
  "Install" documented:
    - the rule;
    - what `@main` and `v0.0.1` give;
    - how to check an installed binary.

    (2026-10-17, https://git.eeqj.de/sneak/sfdupes/issues/116)

- `scan` creates the database's directory where the database path leads. The
  directory is the path with its last element cut off as text, so a `..` in it
  stays. A `..` after a symlinked directory therefore goes through the symlink,
  as it does when the file is opened. With
  `SFDUPES_DATABASE=/srv/link/../new/db.sqlite` and `/srv/link` a symlink to
  `/data/x`, it now creates `/data/new` and the database in it. It used to
  create `/srv/new` and then fail (2026-10-17,
  https://git.eeqj.de/sneak/sfdupes/issues/119)

- `scan` never records its own database. Once the database and its `.lock` file
  are open, it notes the device and inode of each of these:
    - the database file;
    - its `-wal` and `-shm` files;
    - the `.lock` file.

    The walk passes over any file with the device and inode of one of them,
    however it reaches the file. It prints no line for such a file and counts
    nothing. A record an earlier scan left for one of them is deleted
    (2026-10-17, https://git.eeqj.de/sneak/sfdupes/issues/115)

- `scan` never records or reads anything on a filesystem that is not persistent,
  such as `tmpfs` or `proc`. Such a filesystem is one the kernel makes up or
  keeps in memory. `scan` matches them by device ID, however the walk or the
  content phase reaches a path. As the scan starts, it lists the mounted
  filesystems of those types, from `/proc/self/mountinfo` on Linux and from the
  mount list on macOS. No flag turns that off. Below an operand, the walk does
  not enter the default excludes either. So `sfdupes scan /` skips the kernel's
  view of the running system and the temporary directories. On Linux those are:

    ```
    /proc, /sys, /dev, /run, /tmp, /var/tmp
    ```

    On macOS they are:

    ```
    /dev, /private/tmp, /private/var/tmp, /private/var/vm
    /System/Volumes/VM, /System/Volumes/Data
    ```

    A Windows list exists in the code but is not used. An operand at or inside a
    default exclude is still walked. The scan prints a line for each path it
    leaves out, without counting it as skipped. The records there count as
    outside the scanned operands, as for `-x`. `--no-default-excludes` turns off
    the default excludes only, and `scan --help` lists them (2026-10-17,
    https://git.eeqj.de/sneak/sfdupes/issues/124)

- A progress display's rate covers the items completed over the last 5 seconds.
  When stderr is not a terminal, it covers the items since the line before. It
  falls to 0 while nothing completes. A burst after a wait no longer reads as a
  steady rate of over 1,000 files a second. The time remaining still comes from
  the phase's average. Once no item has completed for 10 seconds, the line shows
  `stalled` and where the phase waits in place of the rate, as
  `stalled disk/db`. The places are:
    - `disk`;
    - `db`;
    - `out`, for stdout;
    - `cpu`.

    A phase that may wait in two places names both. Each phase's word is set in
    the code that defines the phase. A line too wide for the terminal loses its
    bar first, then its rate or those words. Four parts stay whole:
    - the phase;
    - the progress;
    - the elapsed time;
    - the ETA, the time remaining.

    The content phase had one display for its checks and one for its reads. Both
    lost those words on an 80-column terminal, at counts like those of README's
    example. That happened for most elapsed times and ETAs in minutes. Every
    other phase kept them whenever the elapsed time and the ETA were both in
    minutes. The content phase writes its last partial batch, the records not
    yet written, under a display of its own (2026-10-16,
    https://git.eeqj.de/sneak/sfdupes/issues/125)

- New subcommand `info` prints the database's details on stdout, one per line as
  a name, a tab and the value:
    - the path and where it came from;
    - the sizes of the database file and of its `-wal` file, when there is one;
    - when the database file was last changed;
    - the schema version;
    - the number of records;
    - how many records have `head` and `tail` hashes;
    - how many records have a content hash;
    - the sizes of all records added up.

    It opens the database as `report` does and shows two phases on stderr,
    `[1/2] open` and `[2/2] count`. The schema version and the counts come from
    one read-only transaction, a read of the database at one moment. The counts
    come from one query (2026-10-16,
    https://git.eeqj.de/sneak/sfdupes/issues/117)

- `report` shows its work on stderr in two phases, `[1/2] open` and
  `[2/2] read`, with the progress display `scan` and `trees` use. The display
  runs from the moment `report` starts until the summary line. `open` says it is
  opening the database, also while it waits on a busy one. `read` counts the
  records, then reads them against that count in signature order. SQLite takes
  that order from the `files_signature` index without sorting, so the first
  record comes at once. `report` forms the duplicate groups itself and holds
  only those of the size it is reading. It writes them ordered by first path
  before it reads the next size. stdout and the summary line are unchanged, and
  SQLite no longer needs a temporary file to sort in (2026-10-16,
  https://git.eeqj.de/sneak/sfdupes/issues/114)

- `trees` shows its work on stderr in four phases, from `[1/4] open` to
  `[4/4] write`, each with the progress display `scan` uses. The display runs
  from the moment `trees` starts until the summary line. `open` says it is
  opening the database, also while it waits on a busy one. `read` counts the
  records, then reads them against that count. `group` counts the directories it
  groups, and `write` the groups it writes. A display's rate names what its
  phase counts:
    - files;
    - records;
    - directories;
    - groups.

    stdout and the summary line are unchanged. When stdout and stderr are the
    same file, as one terminal is, no line of stdout shares a line with the
    display. When they are not, the display keeps being drawn while stdout waits
    on a slow reader (2026-10-16, https://git.eeqj.de/sneak/sfdupes/issues/113)

- Each `scan` phase's progress display names the phase and its place among the
  five, as `[2/5] walk`. It runs from the moment that the phase starts. It
  shows:
    - the progress;
    - the elapsed time;
    - the ETA, the time remaining.

    `load` counts the records first. The walk's total is estimated from the
    records under the operands. The content phase takes its total from SQLite's
    grouping alone. Until the first record is returned, it says it is searching
    the database. On a terminal a phase with an exact total draws its line
    followed by the bar of the progress-bar library. Any other phase draws a
    spinner followed by the line. Every display is drawn again several times a
    second. No line reaches the terminal's last column, however narrow the
    terminal is, even after it is made narrower during a phase. A line too wide
    leaves out its rate, then its bar. The rest stays whole. The line is cut
    short only when the rest alone does not fit, as below about 70 columns late
    in a large scan. When stderr is not a terminal, the line is printed every 5
    seconds whether or not an item completed (2026-10-16,
    https://git.eeqj.de/sneak/sfdupes/issues/112)

- `scan`, `report` and `trees` first print a line on stderr naming the database
  they use, as an absolute path. The line says whether that path is the default
  or comes from `SFDUPES_DATABASE`. `scan` says when the database is new. README
  "Database" says when `scan` writes each file (2026-10-16,
  https://git.eeqj.de/sneak/sfdupes/issues/111)

- The canonical files and model scripts were copied again from `sneak/prompts`
  `next` at `e3f79c4`:
    - `golangci-lint` v2.14.0;
    - lint and test are phases of the `Dockerfile` that write no image, and
      `make test` runs the suite under the race detector. Gone are
      `Dockerfile.lint`, `script/verify-lint-image-pin` and `make test-race`;
    - every `docker build` in `script/` passes a new `CACHEBUST` value. Every
      stage of the `Dockerfile` declares it first, so every `RUN` step runs on
      every build. The steps that compile keep Go's build cache, and the tests
      run with `-count=1`;
    - `prettier` runs on the host, from the `node` and `yarn` that
      `script/bootstrap` installs. The `prettier` and `markdown` stages are
      gone, and the build stage installs `git` and `make` itself;
    - a new push cancels the workflow's older run on the same branch, and a run
      stops after 20 minutes;
    - a coding agent's settings file is deleted.

    Since https://git.eeqj.de/sneak/sfdupes/issues/109 two steps come before
    `ARG CACHEBUST`:
    - the module download in `lint`, `test` and the build stage;
    - the install of `git` and `make` in the build stage.

    Docker keeps them between builds. Every step that checks or compiles the
    code still runs on every build (2026-10-15,
    https://git.eeqj.de/sneak/sfdupes/issues/95)

- With `-x`, an operand under another operand is no longer dropped when a
  directory between the two, the inner one included, is on a different
  filesystem than the outer one. The inner operand is then walked on its own.
  For a regular-file operand the check starts at its directory. The records of
  files gone from below it are deleted (2026-10-14,
  https://git.eeqj.de/sneak/sfdupes/issues/106)

- With `-x`, `scan` keeps the records below a directory on another filesystem
  under its operand, instead of deleting them. It counts their sizes in the size
  census (2026-10-14, https://git.eeqj.de/sneak/sfdupes/issues/99)

- The hash and content phases open each file without blocking and check that it
  is still a regular file. A file that is now a FIFO or device node is warned
  about and skipped instead of hanging the scan (2026-10-14,
  https://git.eeqj.de/sneak/sfdupes/issues/100)

- The scan lock file is named from the database's path with symlinks resolved,
  after `scan` creates the database file if it is missing. A scan through a
  symlink to the database therefore fails while another scan holds the lock
  (2026-10-14, https://git.eeqj.de/sneak/sfdupes/issues/101)

- On a terminal, a warning for the last file of a phase no longer erases the
  phase's finished bar. The bar is drawn again below the warning (2026-10-14,
  https://git.eeqj.de/sneak/sfdupes/issues/102)

- `scan` records the mtime to the nanosecond, as whole seconds in `mtime` plus
  `mtime_nsec`, and compares it at that resolution. A same-size rewrite within
  the same second is hashed again (2026-10-07,
  https://git.eeqj.de/sneak/sfdupes/issues/12)

- The narration was cut from `TODO.md` Completed Steps and from the comments in
  `script/`, `Dockerfile` and `Dockerfile.lint` (gone since
  https://git.eeqj.de/sneak/sfdupes/issues/95). The Workflow section now
  branches from and merges to `next` (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/49)

- `make test-race` ran the test suite under the race detector in a cgo-enabled
  container, outside `make check`. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 `make test` itself runs the suite
  under the race detector, in the `Dockerfile`'s Debian-based `test` phase.
  `make test-race` is gone (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/18)

- A bare `docker build .` failed, naming `script/cibuild` and `script/docker`,
  rather than serve the gates, the lint and test steps, from cache. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 a bare build succeeds, as
  `REPO_POLICIES.md` requires, and may serve the gates from cache. The builds in
  `script/` pass a new `CACHEBUST` value, so theirs always run (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/39)

- `make fmt` and `make fmt-check` run `prettier` over all Markdown, and the
  Gitea workflow checks it. All Markdown was reformatted (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/19)

- `script/lint` writes no image, so a run no longer leaves an untagged one
  behind (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/48)

- Tests cover:
    - a missing database;
    - `scan` keeping stdout empty;
    - its skip warning;
    - the `report` and `trees` summary lines;
    - every subcommand going through `runE`.

    (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/16)

- `.golangci.yml` was replaced with the current canonical copy, which uses
  `gomodguard_v2`, so lint no longer prints a deprecation warning (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/26)

- A test fails when either `hashWorker` cancellation check in `scan.go` is
  removed (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/83)

- A database path holding `?`, `#` or `%` opens exactly the file it names
  (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/55)

- `scan` rejects `--workers` below 1 as a usage error instead of running
  single-threaded (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/10)

- A test fails when either walk cancellation check in `scan.go` is removed
  (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/81)

- A test checks that `scan` refuses a database with another schema version
  (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/64)

- Four inaccurate comments in `cancel_test.go` were corrected, and
  `walkCancelInFlightDirs` was renamed to `walkCancelInFlightFiles` (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/33)

- The `-x` filesystem-boundary rules in `subdirJob` are tested (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/17)

- `scan` creates the schema in one transaction. A version-0 database with a
  `files` table is refused with a clear schema-version error (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/11)

- README documents install, Docker, a daily `cron` scan and how to read and
  check the reports (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/54)

- The `Dockerfile` build stage kept the Go module cache out of `builder`'s home
  and copied the sources with `--chown`, so no `chown -R` walked them. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 there is no `builder` user and
  nothing changes owner. The build stage only compiles, as root, and the tests
  run as `nobody` in the `test` phase (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/43)

- `--version` prints `sfdupes VERSION` to stdout. README documents it and
  `--help` (2026-10-04, https://git.eeqj.de/sneak/sfdupes/issues/15)

- `scan` stops cleanly on `SIGINT` or `SIGTERM`: it commits what it has hashed,
  deletes nothing more, and exits 1 (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/5)

- `report` and `trees` stream the records instead of holding them all in memory.
  The schema gains the `files_signature` index (2026-10-04,
  https://git.eeqj.de/sneak/sfdupes/issues/14)

- Progress prints at once on a non-terminal and uses a real terminal test.
  Warnings print through a spinner instead of racing its redraw (2026-10-03,
  https://git.eeqj.de/sneak/sfdupes/issues/13)

- These operands are warned about and skipped, and the records beneath them are
  kept:
    - symlinks;
    - sockets;
    - FIFOs;
    - device nodes;
    - `.zfs` directories.

    (2026-10-03, https://git.eeqj.de/sneak/sfdupes/issues/9)

- `scan` holds a lock on a lock file beside the database for its whole run, so a
  second `scan` fails at once with exit 1 (2026-10-03,
  https://git.eeqj.de/sneak/sfdupes/issues/53)

- Tests cover stdout write failures in `report` and `trees`. README states that
  `| head` ends `sfdupes` by `SIGPIPE` and that `>&-` writes to `/dev/null`
  (2026-10-03, https://git.eeqj.de/sneak/sfdupes/issues/30)

- `report` and `trees` open the database read-only, and `scan` leaves it out of
  WAL mode, so reading needs only read access (2026-10-03,
  https://git.eeqj.de/sneak/sfdupes/issues/8)

- Four characters are escaped in `report`, `trees` and warning paths:
    - tab;
    - newline;
    - carriage return;
    - backslash.

    The root directory's path is `/` (2026-10-03,
    https://git.eeqj.de/sneak/sfdupes/issues/7)

- A plain `docker build .` stamps the `git` tag or short commit instead of
  `dev`. `.dockerignore` sends `.git` without `.git/config`. The build stage
  stamps the `VERSION` build argument, else `git describe --tags --always`. It
  fails if the context carries `.git` and the version is still empty, `dev` or
  `unknown`. The Gitea workflow checks out the full history (`fetch-depth: 0`),
  so it stamps the same value as `make build` (2026-10-02, branch `next`,
  https://git.eeqj.de/sneak/sfdupes/issues/67)

- The 1 KiB end-window sampling was replaced with the hashes chosen by file
  size, and README "Duplicate detection" documented every size threshold. A file
  under 10 MiB was hashed in full, and its `head`, `tail` and `content` all held
  that hash. A larger file got only its 64 KiB `head` and `tail` in the hash
  phase. The content phase ran after the update phase. It read a file for
  `content` only when its size, `head` and `tail` matched another record's, from
  this scan or an earlier one. `content` was the whole file below 50 MiB, and
  gigabyte-spaced 1 MiB samples at or above. It never read a file gone or
  changed since its record was written. `report` and `trees` left out any record
  without a `content` hash. The `content` column is part of the version 1
  schema. Superseded by https://git.eeqj.de/sneak/sfdupes/issues/132
  (2026-09-22, branch `next`, https://git.eeqj.de/sneak/sfdupes/issues/61)

- The dead `files.dat` references were removed from `Makefile`, `.gitignore` and
  `.dockerignore` (2026-09-21, branch `next`,
  https://git.eeqj.de/sneak/sfdupes/issues/22)

- The lint-image pin comments and `FROM` form in `Dockerfile` and
  `Dockerfile.lint` were fixed. Both pins became the policy
  `# image:vX.Y.Z, YYYY-MM-DD` comment over a bare `FROM image@sha256:...`,
  without the false `(Debian-based)` note or the tag. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 the `Dockerfile`'s `lint` phase
  holds the only `golangci-lint` pin, in that form. `Dockerfile.lint` is gone,
  and so is `script/verify-lint-image-pin`, which compared the two pins
  (2026-08-10, branch `next`, https://git.eeqj.de/sneak/sfdupes/issues/25)

- All linting ran in Docker, via `Dockerfile.lint` and `script/lint`. Per the
  owner ruling the linter is never installed on a host, and
  `golangci-lint config verify` runs before `golangci-lint run`.
  `script/bootstrap` stopped installing or pinning the linter, and
  `script/verify-linter-pin` was retired. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 both commands run in the
  `Dockerfile`'s `lint` phase. `script/lint` builds that phase alone, and the
  build stage depends on it through `COPY --from=lint /src/go.sum /dev/null`.
  `Dockerfile.lint` and `script/verify-lint-image-pin` are gone. Nothing inside
  an image build may run `docker`, so the phase calls `golangci-lint` directly
  (2026-08-10, branch `next`, https://git.eeqj.de/sneak/sfdupes/issues/46)

- The Docker build stage's prerequisites were installed by running
  `script/bootstrap` instead of `apk add --no-cache make` inline.
  `script/verify-linter-pin` failed the build unless the linter copied from the
  lint stage was the version that `script/bootstrap` pinned. Builds then took up
  to 5m14s from an empty cache, mostly in a `chown -R` of the module cache,
  filed as https://git.eeqj.de/sneak/sfdupes/issues/43. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 the build stage installs `git` and
  `make` with `apk add --no-cache` and only compiles. The `lint` and `test`
  phases are the gates (2026-08-09, branch `dockerfile-bootstrap`,
  https://git.eeqj.de/sneak/sfdupes/issues/42)

- The Docker layer cache is defeated for the gate steps, so `script/cibuild` and
  `script/docker` cannot report a pass they did not earn. The `Dockerfile`
  copies the tree before its gates. On an unchanged tree Docker served them from
  cache, and the build exited 0 without running them. The fix was a
  `CHECK_EPOCH` build argument. Since
  https://git.eeqj.de/sneak/sfdupes/issues/95 every `docker build` in `script/`
  passes a new `CACHEBUST` value instead. Every stage of the `Dockerfile`
  declares it before every step that checks or compiles the code. Run as root,
  the tests fail `TestScanHardlinkRunFailsTogether`, because root reads through
  the `chmod(0)` the test relies on. So the `test` phase runs them as `nobody`
  (2026-08-09, branch `cibuild-cache-bust`,
  https://git.eeqj.de/sneak/sfdupes/issues/32)

- `script/bootstrap` checked the installed `golangci-lint` version instead of
  only its presence. The version lived only in `GOLANGCI_LINT_VERSION`, with the
  `go install` module reference derived from it. Any installed version that was
  not the pin was reinstalled, whether:
    - older;
    - newer;
    - absent;
    - unparseable.

    `go install` writes into `GOBIN`, or `GOPATH/bin`, while `make lint` ran the
    first `golangci-lint` on `PATH`. So bootstrap re-read the effective version
    after installing. On a mismatch it printed both paths and both versions and
    exited non-zero. It did not reorder `PATH` or delete anyone's binary. The
    `--version` call kept its stderr and was bounded by `timeout(1)` where that
    exists. `git`, `make` and `go` kept presence-only checks. The version check
    was verified by bootstrapping this host from v2.10.1 to v2.12.2, and again
    to a no-op. Stub runs of the script under `dash` covered:
    - a thirteen-input version-parse matrix;
    - a shadowed install, which must exit non-zero;
    - an install destination not on `PATH`;
    - `GOBIN` set;
    - a wedged binary, which must hit the timeout.

    `make check` and `make lint` were clean at v2.12.2, so v2.10.1 had not been
    hiding findings on `main`. Superseded by
    https://git.eeqj.de/sneak/sfdupes/issues/46, since the linter is no longer
    installed on a host (2026-08-09, branch `bootstrap-version-check`,
    https://git.eeqj.de/sneak/sfdupes/issues/24)

- The hash worker pool unwinds on the error path. The pool is an owned,
  context-aware `hashPool`. Every blocking send in the feeder and the workers
  selects on `ctx.Done()`, and `jobs` is closed on every path out. `hashPhase`
  defers `pool.stop()`, which cancels and then drains `results` until the last
  goroutine has exited. Draining is what frees a worker parked on a send. `ctx`
  is passed from `cmd.Context()` as the first parameter, through:
    - `runScan`;
    - `syncScan`;
    - both worker pools;
    - the whole database layer.

    The walk pool gets the same treatment, plus a `ctx.Err()` guard after the
    walk. A cancelled walk yields a partial size census, and every file that it
    never reached looks vanished to the update phase. That phase's own `BeginTx`
    also fails on the cancelled context before deleting anything. The guard is
    still the barrier once an interrupted scan may commit what it has. Tests
    drive `run(scan)` against a database whose insert trigger aborts. They
    assert that the scan fails instead of hanging, and that the goroutine count
    returns to its pre-scan baseline. Other tests cancel a scan part-way through
    the walk, deterministically, by counting the scan's own checks of
    `ctx.Done()`. They assert that it stops at the guard with every record
    intact. It then holds a partial size census and a still-populated record
    index. Direct tests cover the remaining cancellation branches of both pools.
    They test:
    - `sendEvent`;
    - the walk workers;
    - `dispatchDirs`;
    - `feedHashJobs`;
    - `hashWorker`;
    - `hashPhase`.

    (2026-08-09, branch `hash-pool-cleanup`,
    https://git.eeqj.de/sneak/sfdupes/issues/6)

- The database is closed on every fatal exit path. `fatalf` and its `os.Exit(1)`
  are gone. So the deferred `db.Close()` runs when a subcommand fails, and with
  it the SQLite WAL checkpoint. These return errors instead:
    - `runScan`;
    - `runReport`;
    - `runTrees`;
    - `loadRecords`;
    - `resolveRoots`.

    The single exit point is `run` in `main.go`. It maps a `fatalError`,
    anything a subcommand returned, to exit 1, and cobra's own argument and flag
    errors to exit 2. So a runtime failure is never reported as a usage error,
    nor printed with the usage text. `main_test.go` drives the CLI in-process
    and asserts the exit codes from README "Error handling" and the
    stdout/stderr split. It also checks that a fatal error after the database is
    open leaves no `-wal` or `-shm` file behind, for `scan`, `report` or `trees`
    (2026-08-09, branch `db-close-on-fatal`,
    https://git.eeqj.de/sneak/sfdupes/issues/4)

- `golangci-lint` was updated to v2.12.2 with the canonical config, merged as
  `38a01bd`. The pinned linter in the `Dockerfile` lint stage and
  `script/bootstrap` went from v2.12.1 to v2.12.2, and `.golangci.yml` was
  replaced with the canonical file. Four thresholds now live under
  `linters.settings`, as the v2 schema requires, so they are applied:
    - `lll`;
    - `funlen`;
    - `cyclop`;
    - `dupl`.

    No new lint findings surfaced (2026-08-09, branch `golangci-v2.12.2`,
    https://git.eeqj.de/sneak/sfdupes/issues/3)

- `Makefile` targets were converted to scripts-to-rule-them-all `script/`
  entrypoints, like the other managed repos. All 12 entrypoints exist:
    - `bootstrap`, `setup` and `projectname`;
    - `test`, `lint`, `fmt`, `fmt-check` and `check`;
    - `docker`, `cibuild`, `precommit` and `install-precommit`.

    Every `Makefile` target does nothing but call its script (2026-07-26, commit
    `3abeacf`, https://git.eeqj.de/sneak/sfdupes/issues/1)

- The binary is the default `make` target. Plain `make` builds `sfdupes`, where
  it used to run `check` plus `build`. `make build` remains as an alias
  (2026-07-24, branch `make-default-target`)

- The phases became scan-wide, with concurrent operands and batched updates. All
  operands seed the shared walk pool, and every pass runs once over the whole
  scan. So the totals and ETAs cover the whole scan. The per-operand walk, hash
  and update cycles and their stderr announcements are gone. The update pass
  writes in batches, each in its own transaction. The filesystem is
  authoritative and the database follows it, so one scan's changes need not land
  all at once (2026-07-24, branch `scan-wide-phases`)

- The stat pass was split back out of the walk. Phases are strictly sequential
  again, per operand, with parallelism only inside each phase:
    - walk;
    - stat;
    - hash;
    - update.

    The walk enumerates paths with per-directory workers, and the stat pass
    lstats them with per-file workers. That restores the exact total and ETA of
    the stat bar (2026-07-24, branch `parallel-phases`)

- Each operand is announced on stderr before its passes. With per-operand walk,
  hash and update cycles, a multi-operand run such as `scan /srv/*` showed pass
  totals that looked like the whole run's (2026-07-24, branch
  `scan-operand-progress`)

- The walk became parallel. The walk pass was a single goroutine and took hours
  at about 20 million files on a busy pool. One ZFS server took 4 hours for 22
  million files. It is now a walk by a pool of per-directory workers that
  records size and mtime as it goes. That folds away the separate stat pass and
  halves the metadata reads. Each `PATH` operand commits in its own transaction,
  so an interrupted scan keeps completed operands (2026-07-24, branch
  `parallel-walk`)

- `scan` gained a persistent database. It maintains an SQLite database that
  survives between runs, with the absolute path as each record's key. The driver
  is `modernc.org/sqlite`, pure Go, so cgo stays disabled. A rescan hashes only
  new or changed files, by mtime and size. It deletes the records of files
  vanished from under the scanned operands and leaves records outside them
  untouched. So `scan` can run daily from `cron`. `report` and `trees` read the
  database, with no positional arguments, instead of a scan stream. The database
  is at `/var/lib/sfdupes/db.sqlite`, and `SFDUPES_DATABASE` overrides that. WAL
  journaling plus a single-transaction update keep a report run during a scan
  safe (2026-07-24, branch `persistent-database`)

- The `origin` remote, `git@git.eeqj.de:sneak/sfdupes.git`, was added. `v0.0.1`
  was tagged, and `main` and the tags were pushed (2026-07-23)

- The `scan` command line was reworked. Required `PATH...` operands, parsed by
  cobra, replace the `/srv` `-root` default. The new `-x`, `--one-file-system`
  flag follows the GNU convention and stops at filesystem boundaries, which the
  walk crosses by default (2026-07-23, branch `scan-required-paths`)

- The repo was brought into compliance with `REPO_POLICIES.md` (the version of
  2026-07-06), the existing repo checklist and the Go styleguide. The audit of
  2026-07-22 found these gaps, and the branch closed them:
    - `.gitignore`, with the standard patterns for secrets and for OS, editor
      and Go files. The compiled `sfdupes` binary and `files.dat` had been
      sitting untracked in the tree;
    - `.editorconfig`;
    - `LICENSE` and the README License section. MIT was assumed from house
      convention, for the owner to confirm;
    - `REPO_POLICIES.md` in the repo root;
    - the canonical `.golangci.yml`, with its 150 findings fixed, so that
      `make lint` is clean;
    - the `Makefile` targets the policy requires:
        - `test` and `lint`;
        - `fmt` and `fmt-check`;
        - `docker` and `hooks`.

        `check` depended on `build`, which wrote the binary. `make check` must
        not modify files, so `check` no longer depends on `build`;

    - a test suite, which the policy required with a 30-second timeout and the
      conditional `-v` rerun. At 64% coverage, it covers:
        - parsing and grouping;
        - digests and suppression;
        - hashing and the scan pipeline;
    - a multistage `Dockerfile` and `.dockerignore`. The `Dockerfile` had:
        - hash-pinned images;
        - a lint stage that failed fast;
        - a build stage that ran `make check`;
    - `.gitea/workflows/check.yml`, running `docker build .` on push with the
      checkout action pinned by commit SHA;
    - the README sections the policy requires. The first is a Description whose
      first line names the project, its purpose and category, its license and
      its author. Then come Getting Started and Rationale, and at the end TODO,
      License and Author;
    - the stale README non-goal "no git repository setup and no CI" removed;
    - the pre-commit hook, installed with `make hooks` once that target existed.

    The code was already `gofmt`-clean, so no standalone formatting commit was
    needed (2026-07-23, branch `repo-policy-compliance`)

- `git init` was run, with a README-only first commit. The code baseline was
  committed on `main` (2026-07-22)

- The `scan`, `report` and `trees` subcommands were implemented (pre-git
  history)

# Future Steps

- Possible later features, out of scope per README, are a byte-for-byte check of
  candidates and helper scripts for removal.

# Accepted Divergences

- Flat single-package layout, with the `.go` files in the repo root. That is
  fine for a small single-binary tool per the Go styleguide, and the tracker
  audit agrees.