v0.2.0

Everything a stranger hits in the first ten minutes, every state a person could be stuck in,
and what the app was missing to stand on its own.

Added

• pitboard log shows what pitboard has changed and when, from the record it was already
  keeping. The log now names which front end asked.
• pitboard uninstall deletes every parked login and then pitboard's own files, in that
  order, because the account list is the only index of those keychain items.
• pitboard abandon gives up on an interrupted switch that cannot be finished, keeping
  every login. The way out when recovery cannot reach Anthropic.
• pitboard status --offline answers from what was last measured, without asking Anthropic
  or touching a login. Milliseconds instead of seconds, and it works with no network.
• forget asks before deleting a parked login, unless --yes or --json.
• Each usage row in --json carries severity, Anthropic's own grade for that limit. A
  field added to the v1 envelope; nothing was removed or renamed.
• The app: each account's email, when each limit resets, when a parked login stops working,
  doctor's checks on demand, its own version, a mark in the menu bar, and VoiceOver labels.
  It can record the account in use, drop an account, and run Claude Code's own sign-in for
  a new one, showing what that sign-in says rather than borrowing a terminal.
• cargo binstall pitboard fetches the built binary instead of compiling the tree.

Fixed

• pitboard statusline typed at a prompt waited for input that was never coming. It reads
  stdin only when something is piping into it.
• Offline, the account in use rendered as one with nothing parked, advising a sign-in it did
  not need.
• A switch that failed before installing anything left its journal behind, so the next
  command announced a recovery for something that never happened.
• enroll --sign-in checked what could refuse the enrolment only after the browser sign-in.
• The status line showed ?·? for every account but the one in use until someone ran
  pitboard by hand. It now keeps the numbers Claude Code hands it.
• doctor and forget read pitboard's record of its last switch rather than who is signed in,
  so a sign-in made with Claude Code's own /login made both wrong.
• Three ways a parked login could be left in the keychain with nothing naming it.
• A renewal that could not be written left the account with a login already spent.
• The keychain ceiling has its own error, saying the size, the limit, and what to do. A
  login with MCP server tokens in it is past that limit, which is not theory.
• ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN and CLAUDE_CODE_OAUTH_TOKEN raise a warning
  on every change: Claude Code signs in with those, not with the login pitboard moved.
• Columns line up in what a terminal draws, so a label in Chinese or Japanese no longer
  pushes everything after it out of line.
• One state file serves every credential slot, and what was switched to in one slot is no
  longer claimed in another.
• The menu bar showed the largest percentage of any limit, so a row scoped to one model
  read as though everything had stopped.
• Windows gets one sentence instead of a screen of type errors.

Internal

• MSRV is 1.91, measured by building it, and CI builds at whatever the manifest claims.
• The state file can be read forwards, and says which half to upgrade when it cannot.
• A release is guarded, re-runnable, and carries build provenance; the macOS command line
  binaries are signed and notarised like the app. A tag like v0.2.0-rc1 is a pre-release:
  no crates.io, no update feed.
• The app can be tested without a keychain, and is.