0.1.1

v0.1.1

A patch release: one reported test failure whose root cause turned out to be a class of three,
and two corrections to the media benchmark. No product code changed, so Flea's own field-bench
numbers are untouched.

• backend::kind asserted that /usr/share/mime/application/msword.xml carries the English
  comment "Microsoft Word Document". That file is generated by update-mime-database from
  everything in /usr/share/mime/packages/ and no package owns it, so its untranslated comment is
  whatever an installed application happens to supply: libreoffice.xml gives that string and
  shared-mime-info's own freedesktop.org.xml gives "Word document". The diagnosis in the report
  was distributions shipping different wording, but the reporter runs the same Omarchy 4.0.2-1 and
  the same shared-mime-info 2.5.1-2 as the development box, so the variable is an unrelated
  installed package rather than the distribution. Fixes #1, reported by @tempest-chaoscreator.

• Two more tests in the same class, found by sweeping for it rather than by another report.
  a_type_with_no_untranslated_comment_answers_none passed on the development box only because
  libfm is installed: on a stock box the file it reads does not exist, so it passed by reading
  nothing and proved nothing about the property it names. The thumbnailer tests asserted that
  evince, ffmpegthumbnailer and glycin are installed, and none of those is even an optional
  dependency, so they failed on a stock box for correct behaviour. Content assertions now read
  fixtures through the from_str and from_pairs seams these modules already had, and each
  module keeps one live test that asserts shape rather than vendor wording. The suite is 335
  tests, up from 332, and it passes against a mime root built from freedesktop.org.xml alone,
  which is the root that reproduced the reported failure.

• The media bracket said strata drew no thumbnails and could thumbnail no format at all. Both
  readings came from counting files in the shared thumbnail cache, and strata persists nothing
  there: it renders each thumbnail into a scratch directory it deletes on drop and holds the image
  in an in-memory cache. One blind instrument produced two agreeing wrong numbers. Reported on the
  release thread by the project's operator, who had seen the thumbnails on screen.

• strata is now measured by a live watch and ranked. It draws 205 thumbnails on the media
  fixture, settles in about 30,800 ms, and uses 70.3 MiB and 1.82 s of CPU. Its work count and its
  timing come from the same runs rather than from two instruments on two days. Flea keeps both
  columns it wins, settled listing first of six and CPU first of seven, though the CPU lead is now
  1.28x over strata rather than 3.61x over nemo; its memory placement moves from fifth of six to
  sixth of seven, because strata uses less. The 100,000 file bracket is unchanged.

Timing figures are medians of three cold runs and should be read as magnitudes rather than
constants: repeat batches on the same box disagree by a factor of two or more. The CSVs, their
manifests and the chart generator ship in docs/bench/, so every number above can be re-derived
and every committed chart re-checked.