v0.2.0

Install

cargo install mandible

Or download a binary below. Verify with the accompanying .sha256.

---

Six fixes, three of them found by rendering a deliberately awkward CLI through a
real pseudo-terminal rather than by any test.

Fixed

• The USAGE line no longer repeats the command name. docker import
  rendered as import docker import [OPTIONS] file|URL|- [REPOSITORY[:TAG]].
  The check asked whether the usage's first word was the node's name, but
  cobra and argparse both print the full command path, so the name was stapled
  on the front of nearly every subcommand of every such tool. It now scans the
  whole leading run of command words. Tools that print no name at all
  (Usage: [OPTIONS] FILE) still get one added, which is what the prepending
  was for.

• A token wider than the pane is broken across lines instead of discarded.
  It was ellipsis-truncated, so a 150-character URL rendered as
  https://registry.example.com/v2/org… with everything after it unrecoverable
  from the parsed view. Splits are placed by display width, so a double-width
  character cannot straddle the boundary and overflow the pane.

• A relative tool path works. mandible ./scripts/tool.py failed with
  "No such file or directory" for a file plainly present: the path was checked
  against the caller's working directory, then the probe ran with its own
  directory redirected into a scratch dir (§6 rule 8). Resolution now yields an
  absolute path — via std::path::absolute, deliberately not
  fs::canonicalize; see below.

• argparse subcommands survive a styled section heading.
  add_subparsers(title="commands") is the ordinary way to name that block,
  and the dedicated scan was gated on the heading reading positional arguments, so a styled heading collapsed the entire command tree to a single
  node. The scan's structural evidence — a {a,b,c} pseudo-entry with deeper
  lines beneath it — is stronger than the heading text ever was, and still
  refuses a plain positional carrying choices=[...].

• A command list at the same indent as its heading is recognized. dnf 4
  prints its whole command list flush at column 0 under a flush-left heading;
  the engine required content indented more than its heading, so mandible dnf showed one node and no subcommands. Now 30.

• A pending row's spinner no longer touches the name. dnf's longest
  command rendered check-update⋯ loading, one mangled word rather than a name
  and its status — the same defect fixed for summaries in an earlier release
  (apt-get's dselect-upgradeFollow) and missed in the sibling branch,
  because no tool in the suite had a pending row at the column until dnf
  gained subcommands.

Notes on two near-misses

Both were caught by the PATH-wide coverage sweep and by nothing else; the unit
suite was green through both.

• Making resolved paths absolute with fs::canonicalize defeated §6 rule 0.
  is_never_probe matches on the file name, and reboot, poweroff,
  shutdown and telinit are symlinks to systemctl — resolving renamed them
  before the refusal ran. It also broke ten iptables* tools, which dispatch on
  argv[0]. Fixed by using std::path::absolute, which does not follow links.
• The same-indent command rule initially fabricated 28 subcommands out of
  mysqlslap's config-variable table (port 3306, no-drop FALSE), because at
  a shared indent every row is a candidate heading for the rows beneath it and
  init-command contains the word "command". A heading must now not itself look
  like a row.

Final sweep is identical to baseline on every aggregate — 89.19% described
across 2266 tools, 1 suspicious, 320 verbatim — with zero status changes, zero
nodes lost, and dnf the only gain.

Internal

• scripts/smoke_cli.py: a deliberately awkward argparse CLI for exercising
  layout by hand — a twelve-level command chain, four flag-table shapes, tokens
  with no whitespace to wrap at, and sixty flags. It found three of the bugs
  above within minutes of existing.