longecho 0.5.0

Changed

• Behavior change: longecho build now overwrites a non-empty site/
  only when its README.md carries a generator starting with
  "longecho build". A site/ with another tool's stamp, a README without
  one, or no README at all (for example a memex HTML viewer) is left alone
  unless you pass --force. Previously only sites with a foreign stamp were
  protected, so unstamped viewers were silently overwritten.
• README titles and descriptions are read from the parsed Markdown instead
  of a line scan. Descriptions no longer start with HTML comments, raw HTML
  blocks, or badge markup; lists such as a table of contents no longer count
  as the first paragraph; setext headings count as titles; inline markup
  becomes plain text. This renders each README, so query over a large tree
  is slower (0.8s to 2.0s for 269 sources in testing).
• The version now lives only in src/longecho/__init__.py; pyproject.toml
  reads it through [tool.hatch.version].

Fixed

• longecho build --output pointing at an existing file is refused instead
  of crashing.

Added

• Documentation site at https://queelius.github.io/longecho/ (Material for
  MkDocs, deployed from main), and a Documentation project URL.
• .gz and .tgz accepted as durable archive formats. Gzip has shipped with
  every mainstream OS since 1992 and decodes without special tooling.
  Compound extensions like .jsonl.gz, .csv.gz, and .tar.gz qualify via
  their terminal .gz suffix.
• Compliance smoke tests for .jsonl.gz and .tar.gz pin the behavior and
  document the contract (there is no separate .tar.gz entry; the .gz
  suffix is what earns durability).

Changed

• README durable-formats section clarifies that gzipped files qualify via
  terminal .gz, covering .jsonl.gz, .csv.gz, .tar.gz, etc.