v6.6.0

Fixes

• Files are now replaced atomically: a crash, kill, or full disk mid-write can
  no longer destroy the original image.
• Archive members whose optimization fails are kept in the repacked archive;
  previously they were silently dropped.
• Nested archives (e.g. a CBZ inside a ZIP) now optimize; previously they always
  errored.
• Archives keep their comments and member order through repack; tar symlinks and
  hardlinks survive; 7z member times are preserved; pre-1980 member times no
  longer crash conversion; legacy non-UTF8 zip member names are repaired.
• Archive conversions compress each member by content; converting a rar, tar, or
  7z no longer leaves every member stored uncompressed.
• Files that fail to optimize are retried on the next timestamped run;
  previously directory timestamps covered them and they were skipped forever.
• One corrupt file inside an archive (e.g. a decompression bomb) no longer
  aborts the entire run.
• Boolean options set in config files or environment variables are no longer
  ignored.
• Requested archive conversions (e.g. CBR to CBZ) now happen even when the
  contents are already optimized.
• Lossless WebPs with large metadata inside archives are no longer misdetected
  as lossy.
• Owner-password-restricted PDFs keep their encryption; optimized PDF images
  keep their color transform; more signed PDFs are detected and refused.
• svgo no longer strips viewBox, and keeps title/desc metadata unless
  --strip-metadata is given.
• Conversions discarded for being bigger (and dry runs) are reported as skips,
  not errors.
• --preserve no longer fails as a non-root user when it cannot change
  ownership; it still restores permissions and modification time.
• Temporary files and animated-WebP frame directories are cleaned up when a tool
  fails or a run is cancelled, instead of leaking to disk.
• picopt doctor no longer reports missing tools and exits nonzero on healthy
  installs.
• Symlink loops, unreadable directories, and FIFOs no longer hang or crash the
  walk.
• Bad --memory-limit and --after values abort with clean error messages;
  timezone-aware --after values are honored; lowercase format lists are
  accepted for -x and -c.
• -I/--no-default-ignores without -i no longer aborts at startup.
• Existing timestamps are invalidated once after upgrading because the timestamp
  config check gained new keys; the first -t run re-examines already-optimized
  trees.

Features

• Per-directory .picopt.yaml config files: pin any setting to a directory
  tree. Deeper directories win; environment variables and the command line still
  override. Editing one re-processes its tree on timestamped runs.
• Write your invoked options to config files: -w to the user config, -W to a
  .picopt.yaml in each target directory, --write-config-file PATH to an
  explicit path.
• Memory-aware scheduling: picopt now estimates the peak memory of large
  archives and limits how many run at once so big libraries (e.g. multi-GB comic
  archives) no longer exhaust RAM and get the process OOM-killed. Tune the
  budget with the new --memory-limit option (e.g. --memory-limit 8G), which
  reads as an approximate peak-memory target; the default is two-thirds of
  detected RAM.
• Incremental archive re-optimization: when a timestamped archive changes,
  members older than its timestamp are skipped instead of re-optimized; -E
  disables the member check. Timestamp files found inside archives are consumed
  and removed on repack.

Performance

• Archive member format detection runs in parallel workers instead of
  serializing the scheduler.
• Solid 7z archives extract in one pass instead of decompressing from the start
  for every member.
• The progress pre-scan is skipped in quiet mode, less image data is shipped
  between processes, and scheduling under a memory limit no longer rescans the
  whole queue on every completion.