v0.6.0

The crate moves to Rust edition 2024, the configuration file finds itself, the
gui backend is brought much closer to web, --theme finally reaches every
backend, and the backends are renamed.

Changed

• Rust edition 2024, with 1.95 as the minimum supported version — the
  highest rust-version declared anywhere in the tree, and confirmed by
  building and testing on a pinned 1.95 toolchain rather than by reading
  manifests. Many crates declare none at all, which is why the build is the
  proof and the declaration only the starting point.

• Dependencies brought up to date: comrak 0.55, the egui stack at 0.36, wry
  0.57 and tao 0.37, resvg and usvg 0.48, mermaid-rs-renderer 0.3, base64 0.23.
  comrak's syntect-onig feature is now selected explicitly: from 0.53 comrak
  stopped enabling a regex engine on its own, which would have broken a
  webview-only build.

• A selected set of clippy lints is enforced in Cargo.toml. They were all
  fixed across the crate first, so the list is a floor the CI holds rather than
  a wish. clippy::pedantic is deliberately not switched on in bulk: the CI
  lints on a floating stable toolchain, and a lint added upstream would break
  unrelated pull requests.

• The config file is written on first run, and its location is resolved
  rather than hard-coded. In order, on every platform: an existing
  ~/.config/mdr/config.kdl, then $XDG_CONFIG_HOME/mdr/config.kdl when that
  variable holds an absolute path, then %APPDATA%\mdr\config.kdl on Windows,
  then ~/.config/mdr/config.kdl. %USERPROFILE% gives the home directory on
  Windows, because Git Bash sets HOME to a POSIX path a native binary cannot
  resolve.

  Putting an existing ~/.config/mdr/config.kdl ahead of XDG_CONFIG_HOME is
  a deliberate departure from the spec, which says the variable wins: every mdr
  before this release read that one path and nothing else, so deferring to the
  variable would silently ignore the config of everyone who has both. A
  relative XDG_CONFIG_HOME is ignored with a warning, as the spec requires. Being unable to write
  the file is a warning, not a failure: mdr runs on the defaults the file would
  have carried anyway.

  The generated file selects backend auto, which every build has — it used to
  hard-code webview, which a --no-default-features --features tui-backend
  binary cannot run. When nothing names a home directory mdr still reads
  ./.config/mdr/config.kdl if it is there — and treats it like any other
  config, including correcting an old backend name in it — but does not create
  one, so no .config/ is left behind in the directory it was started from.

• The backends are named gui, tui and web. egui and
  webview said which library draws the window, which is not what someone
  choosing a backend is picking between: a native window, a terminal, or the
  system webview. mdr --backend web, backend gui in config.kdl.

  A config file written before this release is corrected in place the first
  time it is read: backend egui becomes backend gui, comments and every
  other setting are preserved, and the run says so once. A file mdr cannot
  write — read-only, or a symlink — is left as it is, with a warning, and the
  run carries on. On the command line there is no such mapping: --backend egui is simply not a
  backend any more, and the error lists the names that are.

  The Cargo features keep the crate names (egui-backend, webview-backend,
  tui-backend), so documented build commands still work.

Added

• A LICENSE file. MIT was declared in Cargo.toml, in the README, in the
  Homebrew formula, in the Scoop manifest and in the Chocolatey licenseUrl —
  which pointed at a 404 — but the text was nowhere in the repository, so
  GitHub reported no licence at all. Every packaging recipe is now configured
  to install it: the .deb and .rpm, the release archives (which is where
  Homebrew and the AUR package pick it up), and the Nix build. None of those
  formats has been built from this branch.

• -s, --set-default-backend <BACKEND> writes the backend into the config
  file and exits. Only that value is replaced — comments and every other
  setting are preserved — and mdr refuses to write a backend this binary was
  not built with, which would only fail on the next start.

• The publishing setup documents every channel the release workflow drives:
  PACKAGING.md named four of the seven secrets and three of the six
  variables, leaving Chocolatey, Scoop and the Snap Store undocumented even
  though their jobs run. The README gained the cargo install mdr /
  cargo binstall mdr route, which was never written down.

• Short forms for the options that lacked them: -t for --theme, -c for
  --config and -l for --list-backends. --theme was also missing from the
  man page's option list, though it was already documented as a config key.

• The gui and web backends share one palette, one body size and one h1
  (src/core/style.rs): a 16 px body, headings at 2 / 1.5 / 1.25 / 1 / 0.875 /
  0.85 em, code at 85 %, GitHub colours. gui used to run on egui's defaults —
  a 13 pt body with an 18 pt heading, so h1 through h6 all landed within
  five points of each other — and web rendered h3 to h6 at the browser's
  own sizes, which were a third scale again. The stylesheet is generated from those
  constants, so the palette and the body size stay in step.

  h2 to h6 are the stylesheet's in web only. gui renders through
  egui_commonmark, which interpolates its own sizes between the heading and
  body styles; the API it exposes today takes those two ends and no table, so
  setting them makes h1 and the body agree while the levels between come out
  larger.

  Thirteen named pairs from this palette — the document, the code backgrounds
  and the sidebar states — are checked against the WCAG 2.2 contrast minimum by
  unit tests, including muted text on a hovered sidebar entry, which
  GitHub's own #656d76 misses at 4.4999:1, so the light palette uses #646c75.
  egui's default inline-code chip put body text at 3.1:1.

• gui picks the platform's UI and monospace faces (SF Pro / SF Mono,
  Segoe UI / Cascadia Mono, Cantarell / Noto Sans Mono…), matching the
  system-ui and ui-monospace stacks the stylesheet asks for, plus a short,
  bounded list of fallbacks for the scripts those two do not cover.

• **bold** is visible in gui again. egui has no font weights —
  strong() only changes the colour — so bold and body text were drawn
  identically. Strong text has its own palette entry now, with a test that they
  cannot be equal.

• The gui table of contents follows the web sidebar: a small uppercase
  muted label instead of a document-sized heading, and entries graded by depth.
  Its scroll area no longer shrinks to its widest entry, which is what put
  egui's floating scrollbar on top of the text (the rest of #27).

Fixed

• --theme dark|light only ever reached the terminal backend. It was
  accepted everywhere and applied nowhere else: gui and web followed the
  desktop's colour scheme whatever the flag said, so --theme light on a dark
  desktop rendered dark. Both now honour it, and auto still asks the
  environment. The toggle — see below — continues to flip whatever the window
  is currently showing, forced theme included.

  Mermaid is the exception. A diagram rendered natively to SVG carries its own
  colours, whatever the setting; only the web fallback, drawn by the bundled
  Mermaid library, follows it, and only when the page loads — one already on
  screen is not recoloured by the toggle.

• The theme toggle is now t, and works in all three backends. It was
  Ctrl/Cmd+D in web only — the combination Ghostty, iTerm2 and Terminal.app
  bind to splitting a pane, and one the other two backends had no equivalent
  for. t carries no modifier, which is why it was chosen.

  The terminal's bottom bar offers it too. That bar was one fixed string clipped
  to the width it was given, which on an 80-column terminal cut it mid-item and
  hid everything past the search hint; it now drops whole hints instead, keeping
  the most useful ones.

• A light syntax theme was barely legible in the terminal. syntect picks its
  foregrounds for its own background, and the terminal's is whatever the reader
  set, so --theme light drew dark text straight onto a dark terminal. Code
  blocks now paint the theme's background and pad to the frame width, which
  makes the block self-contained — the same thing gui and web do with their
  code background.

• The gui backend read every installed font into memory. It loaded the
  whole of each font file, once per face that file contained, and pushed all of
  them into both fallback chains — 890 faces across 473 files on the machine
  this was measured on, 4.57 GB of font data retained before egui had built
  anything out of it, on a document of any size. Another machine has another
  font collection; the shape of the problem carries over, the number does not. Faces are now chosen from the
  font database's metadata and only the handful selected is read: the body font,
  the code font, and a short list of fallbacks for scripts those two do not
  cover, under a budget of 64 MB and eight faces. The same document now holds
  three system faces and 13 MB.

  mdr --verbose reports what was loaded and what it cost.

• A document opened wherever the last one had been left. eframe's
  persistence feature restores egui's memory, and a scroll area's offset is
  part of it, so the gui window came up scrolled — on this repository's README
  that is 3630 points down, past the logo and the title, which reads as the top
  of the document simply being absent. Every document now opens at its top.

• gui came closer to web. A rule under h1 and h2, as the stylesheet
  draws; code blocks highlighted with the same syntax themes as the terminal;
  a table of contents graded by size and indent instead of by an accident of
  styling — RichText::strong() sets a colour of its own, so the top two levels
  rendered as plain text and the third as a blue link, in one list, while the
  deepest stay deliberately muted — and long entries wrap instead of being cut
  to an ellipsis.

• gui printed raw HTML at the reader. web hands HTML to a real engine;
  gui has none, and egui_commonmark passes an HTML block through as text, so
  a README that centres its logo and title with <p> and <h1> showed its own
  markup. A short, explicit set of tags — headings, paragraphs, images — is now
  rewritten as the Markdown that means the same thing, so it travels the
  pipeline that was already there: the image paths are resolved, the file is
  validated and the SVG rasterised exactly as for ![](…), and a heading is
  listed in the table of contents. A declared width is honoured for vector
  images. Anything outside that set keeps its text and loses its tags, and that
  text is escaped, so a * or a # an author wrote inside a tag stays the
  character it was.

  Only blocks at the top level are converted: the replacement works on whole
  lines and cannot carry back the > of a quote or the indent of a list item.
  This is not an HTML engine — align="center" has no Markdown equivalent and
  is dropped, <br> becomes a space, so the content comes back but its layout
  does not. A heading written in HTML therefore appears in the gui table of
  contents, where web does not list it.

• A refused image was still read from disk in gui. An image outside the
  project root was left as it was written rather than inlined, which is not the
  same as refusing it: egui_commonmark enables egui_extras' file loader,
  which gives a schemeless destination a file:// prefix and reads it, without
  passing through mdr's resolver at all. Neither that loader nor the HTTP one is
  built in any more — the decoder is pulled in without them — so no image is
  read from disk or fetched behind mdr's back: a local or remote one reaches the
  page only through mdr's own resolver, and a data: URI written into the
  document is still shown as the author wrote it. A refusal also replaces the
  image with a note, so the path is no longer something a loader could read.

  For the same reason, the four SVG rasterisers now share one set of usvg
  options that refuses an xlink:href pointing at a file. Checking the drawing
  mdr was asked for said nothing about the files that drawing referenced. (They
  also had four copies of the system font database between them; there is one.)

  The SVG decoder is gone from the window too. It built its own options, so it
  had none of this — mdr rasterises every drawing itself instead, including the
  ones that arrive as data: URIs, which is what a remote badge usually is.

• Searching for "Copy" found one match per code block. The button sits
  inside the content so it can be positioned over its block, and the search
  walked every text node there — so mdr's own control turned up as if it were
  part of the document, highlight included. The walker skips it now.

• The copy button said "Copied" when nothing had been copied. Where the
  asynchronous clipboard is unavailable, the fallback goes through
  execCommand, which reports a refusal by returning false rather than by
  throwing — so the try around it caught nothing and the button claimed
  success anyway. All three outcomes now reach the state that matches them, and
  the hidden textarea is removed even when the attempt raises.

• --offline did not hold in the terminal. The README and the man page
  promise that it stops every network access; tui fetched a remote image
  anyway, because core::offline was not even compiled for it. It now refuses,
  and a test counts the calls to prove none is made.

• A piped document could not find its own images. cat README.md | mdr
  writes the document to a temp file, and every backend then looked for
  assets/logo.svg next to that — so a piped README showed a broken image
  where its logo should be. Relative image paths now resolve from the
  directory mdr was run in. That is a convention, not a deduction: nothing tells
  mdr where the file handed to cat came from.

• gui opened at whatever size the last document had been left at. eframe's
  persistence feature stores the window geometry and egui's memory in one file
  shared by every document, and mdr never asked for any of it — MdrApp has no
  save. It also explains the two defects above it: the restored scroll offset,
  and a theme preference outliving the run that set it. The feature is off, and
  gui and web now both open at the 1100x900 they have always declared.

• Tables in gui were not laid out as tables. The viewer draws one as a
  striped grid inside a frame: no cell borders, no padding, and a header row
  drawn exactly like any other, with each box sized to its own content rather
  than to its column. They are now measured and drawn here — bordered, padded
  cells on a fixed column grid, every cell of a row the height of the tallest,
  with the header set apart by its background and the delimiter row's alignments
  honoured. A cell keeps its inline code, emphasis and links, read from the same
  comrak parse as the rest of the document rather than from the pipe syntax
  again.

• The document column is capped at 900 points in gui, the width the
  stylesheet gives web. Prose used to run the full width of the window and a
  code block was stretched to it.

• web code blocks have a copy button, which gui has always had. It
  survives a live reload, and falls back to a selection where the asynchronous
  clipboard is unavailable.

• Every document in web opened with an empty band above it. The content
  box has padding, so the first element's own top margin could not collapse into
  it and was added to it instead.

• A .webp file was accepted on its container alone. RIFF also begins a
  WAV and an AVI; the check for the WEBP signature that follows it sat inside
  a branch reached only when the file did not start with RIFF, so it never
  ran.

• Following a link in web left the previous document's watcher running.
  The watcher was leaked so that it would outlive the call that started it, so
  every document opened added an OS watch that nothing could stop; a long
  browsing session accumulated them. watch_file now hands the caller a Watch
  that owns the watcher, and the web backend assigns over it — which drops the
  one it is leaving.

• A terminal reporting no rows made the terminal backend panic, on the
  subtraction that places the bottom bar. script -q /dev/null mdr --backend tui is one way to get there.

• A failure between entering raw mode and the cleanup left the shell raw,
  and on the alternate screen. The restore now happens on every way out,
  including a panic.

• cat doc.md | mdr --backend tui is usable again. The document was drawn
  and the first key press then killed it with "Failed to initialize input
  reader", leaving the terminal on the alternate screen.

  The cause is specific to macOS. With the document on stdin, crossterm falls
  back to /dev/tty for the keyboard — but that is a clone device, and the
  kernel refuses to register it with kqueue: EVFILT_READ returns EINVAL.
  mio's registration fails, the event source is never built, and crossterm
  swallows the error until the first read. mdr now reopens the real terminal
  device (/dev/ttys004 rather than /dev/tty) onto stdin before starting, so
  crossterm takes its ordinary path.

• A document taller than 65535 rows no longer wraps around. An element's
  height was narrowed to u16 while a wrapped paragraph is as tall as its line
  count, which nothing bounds to the terminal. Everything derived from it — the
  document height, search offsets, scrolling — went wrong past that point; a
  65545-row document reported 9.

• The terminal backend caps the SVGs it rasterises, like the three other
  rendering paths already did. The size declared by the document decided the
  buffer to allocate outright, so an SVG claiming 40000x20000 asked for one that
  size. It is now scaled down to fit 8192 per side, never up.

• --help now lists every value the parsers accept. --theme named none at
  all, and --backend left out auto — the one the generated config file
  selects. A test ties both lists to what the parsers take, so they cannot
  drift apart again.

Tests

• The file watcher — live reload, used by all three backends — had no tests at
  all. It now has five, driven through real edits and through the write-then-
  rename an editor performs, with a neighbouring file that must stay ignored and
  a watch replaced by another the way following a link replaces it.
• The terminal backend is driven through a real pseudo-terminal, which is the
  only way to cover starting it, pressing a key and leaving cleanly; that is
  what catches the piped-document defect above.
• --set-default-backend and the config resolution are covered through the
  command line, not only through their functions. The WebP signature check is
  covered where it lives, by calling the validator with a RIFF container that
  is a WAV and an AVI.
• The tables and the HTML gui renders itself are covered against what they
  replaced: a _ inside a word, an escaped *, a link whose destination holds
  parentheses, an empty cell at the edge of a row, the delimiter row's
  alignments, and — for HTML — text that would otherwise become a heading, a
  list or emphasis on the second reading.
• --offline is shown to make no request at all, by counting the calls to an
  injected fetch rather than by pointing at a URL that would fail anyway.

Removed

• --init. The config file is created at startup, so the flag had nothing
  left to do. A file the program can write for itself is not worth an option,
  and needing to run it first was the only thing keeping new users from
  discovering that mdr is configurable at all.

Full Changelog: v0.5.1...v0.6.0