v0.2.1

Correctness patch. Every entry below was found by probing the published
0.2.0 rather than reading the source, and each one is a case where the
harness quietly did the wrong thing, panicked inside a dependency, or —
in the worst of them — hung itself.

Changed

• paste now transforms the text the way a real terminal does, so what
  the application receives matches what a user pasting would produce.
  Line breaks become \r (the byte Enter sends; applications in raw
  mode never see \n from a terminal), and while bracketed paste is
  active, paste markers embedded in the text are removed — previously an
  ESC[201~ inside the text ended the paste early and the remainder
  arrived as ordinary key presses. send_str remains the untransformed
  path.

Fixed

• The harness can no longer deadlock itself. Query replies were
  written by the reader thread, so an application that emitted queries
  faster than it read the answers filled the PTY's input queue, blocked
  that write, and stopped the drain — after which the child blocked
  writing and neither side could proceed. Reproduced in the default
  configuration with no test input at all: the wait timed out with a
  stale screen and then Drop never returned. Replies now go to a
  dedicated responder thread, so the drain never writes; undeliverable
  replies are counted and reported ("the application is not reading its
  input") instead of stalling anything. Drop's reap is bounded too —
  teardown must always terminate.
• Mouse reports now follow the UTF-8 encoding (mode 1005) when the
  application selects it. The encoding was collapsed to "SGR or not", so
  a 1005 application received the legacy form — identical below column
  95, and a bare non-UTF-8 byte past it, which such an application
  cannot decode.
• The unanswered-query diagnosis no longer misattributes unrelated
  failures. It was recorded once and never cleared, so a single
  deliberately-unanswered probe at startup (kitty's CSI ? u is the
  common one) claimed to be the cause of every later timeout. A query is
  now only blamed while the application has produced no output since
  asking; otherwise it is reported as context. Every unanswered query is
  named rather than just the most recent, the set is bounded, and
  wait_frame and the Eof errors carry the note too — previously
  wait_frame withheld it, which is the worst place for it to be
  missing, since an application blocked on a probe never reaches its
  first repaint.
• wait_frame timeouts embed the live screen, like every other
  wait. They previously embedded the last completed frame — which can be
  arbitrarily old — under a header saying "screen at timeout", so the
  one place a CI log is the only evidence showed the wrong screen. The
  count of observed frames is still in the message.
• A zero terminal dimension is now a typed Error::Input from spawn
  and resize instead of a panic inside the emulator. In release builds
  — the profile the stress workflow uses — it was worse than a panic:
  the arithmetic wrapped, both calls returned Ok, and the emulator
  panicked on the reader thread, silently killing the drain so every
  later snapshot was blank and a careless test went green.
• current_dir pointing at a path that is not an existing directory now
  fails the spawn instead of being silently ignored: the PTY layer
  falls back to the home directory, so a directory-sensitive test could
  pass against the wrong tree with no error anywhere.
• spawn("") now fails with a one-line Error::Spawn naming the problem
  instead of surfacing the PTY layer's entire PATH search. Genuine
  "program not found" failures keep their underlying diagnosis.