fix: survive the instant-exit pty race; report signal deaths

[vyncint]

The stress workflow caught the suite flaking at ~1/80 instant-exit
spawns (child writes and exits within its first millisecond): the OS
pty teardown — macOS especially — can discard output still buffered at
slave close, and can surface teardown as a signal-death instead of the
real exit code. Reproduced locally at ~1/10 full-suite runs.

Harness fixes:

• attach the reader thread BEFORE spawning the child, so the drain is
  running from the child's first byte (narrows the race as far as
  userspace can; the residual kernel-side caveat is documented in
  docs/DESIGN.md and README known-limitations)
• ExitStatus now carries the terminating signal's name; Display says
  'killed by signal: ...' so harness-level kills can never be mistaken
  for application exit codes (ExitStatus: Clone, no longer Copy)
• Screen's Debug is the compact header+text rendering; a failing
  Result test prints a readable screen instead of a single-line dump
  of thousands of cells (long enough that CI log pipelines dropped it,
  which is why the first stress failures were undiagnosable)

Suite fixes:

• every instant-exit script now ends with a stdin 'read' guard and is
  released with Enter only after its output is asserted — the
  deterministic pattern, now documented in DESIGN.md; doctests teach it
• unicode-torture gained the same guard
• resize-echo initializes crossterm's event source (the SIGWINCH
  listener) before its first draw, closing a resize-before-listener
  lost-signal race
• fixture_bin() always refreshes fixture binaries once per process:
  cargo test does not relink plain bin artifacts, so an old
  target// from a previous cargo build silently tested
  stale fixture code
• exit-code assertions include the full status in their message

Validation: 30/30 full-suite release iterations at --test-threads=4
clean locally (previous flake rate ~1/10); signal reporting covered by
a new signal-death test.

Signed-off-by: Vyncint Ng vyncint@icloud.com