Audio Analyzer 0.15.3

Better detection of fake lossless files, and a website.

Changed

• Low-bitrate fakes in 24-bit files are caught. A 64 kbps MP3 decoded to a 24-bit/48 kHz FLAC was called
  Lossless with high confidence (found while making the promo video), and a 96 kbps one only Inconclusive. Two causes,
  two fixes:

  • The floor that "empty" is measured against was 24-bit rounding noise. A lossy decoder leaves residue far below
    any recorded sound but far above that, so the band above the cutoff looked like content up to 24 kHz. The floor
    of 24-bit files is now taken at 20 bits, about the best real recording equipment achieves.
  • A low-bitrate encoder moves its cutoff from moment to moment, so the typical spectrum only slopes. The loudest
    moments still stop at a hard ceiling with nothing above it. A ceiling between 11 and 19 kHz, at least 40 dB
    deep and with the band above it empty, now counts as the encoder's cutoff ("Hard ceiling at ... kHz").

  None of 289 CD-verified tracks shows such a ceiling. In the 440-file golden snapshot one verdict changed: "The
  Third Man Theme" (a 24/96 file of a 1949 recording) from Possibly lossy (Inconclusive) to Lossy, medium confidence;
  with the realistic floor its content ends near 19 kHz and empty slices show. Seven hi-res files report their
  content ending at 92-96 kHz instead of exactly at the top, and one 96 kHz master's 46.3 kHz drop is now called its
  normal anti-aliasing filter. Library results are redone on the next scan (analysis version 3).

• A steep drop where encoders put their cutoff, whose top band moves slightly too much to count as an encoder's on
  its own, is now called Lossy (medium confidence) when a few high-frequency slices are also empty. Each clue alone
  stays unresolved; no known-lossless reference shows both. This catches Vorbis q3 as decoded by Ubuntu's ffmpeg,
  which was only a warning (the CI test no longer expects a failure). Of 440 real library files, 3 changed, all from
  Inconclusive to Lossy: tracks of the CD-verified "Bartender Mix" compilation, whose other tracks were already
  Lossy. The golden snapshot baseline and the verdict regression cases are re-recorded.

Added

• A product website, https://audioanalyzer.pages.dev, and a proper download page on GitHub. Both link to the
  latest installer, so they need no change when a release comes out.