v1.2.0

Printer reliability and preview accuracy. calibrate() no longer
mis-registers labels when the printer is busy, send() can be cancelled and no
longer exhausts the thread pool under a large fan-out, and the renderer now
previews ^SN serial numbers and scopes fonts per field the way a printer does.
Every renderer rule was checked against Labelary, and the printer behavior was
measured on a GX420t (see HARDWARE-VALIDATION.md).

One source-breaking addition: PrinterError gains a printerBusy case. A
switch over PrinterError with no default needs a new arm. Everything else
is additive.

Added

• ZPLPrinter.maxConcurrentSends (8): at most this many send() calls run
  at once per process. Further sends wait for a free slot, without holding a
  thread, before connecting. Each in-flight send blocks a GCD thread for its
  connect and write, so a large fan-out to unreachable printers could
  previously exhaust the thread pool and stall every other send and unrelated
  global-queue work.
• ZPLPrinter.waitUntilIdle(timeout:): polls ~HS until no formats are
  queued, no partial format is in progress, and no batch is printing (two
  consecutive idle readings). Throws the new PrinterError.printerBusy if the
  printer is still busy at the timeout.

Fixed

• calibrate() (and so setup()) could stall queued labels and print them
  out of register. Tilde commands run the moment they arrive, ahead of queued
  formats. Measured on a GX420t (V56.17.17Z): with two labels queued, ~JC
  calibrated first and the two labels did not print at all until the next job
  arrived minutes later, and then printed out of register (~HS reported no
  formats queued while they were stuck). This matches the
  misaligned labels previously seen after setup(). calibrate() and
  calibrateFull() now wait for the printer to go idle first (idleTimeout,
  default 60s). With the fix, calibrate() held the ~JC for 2.7s and the
  calibration ran only after every queued label had printed.
• send() now observes task cancellation once it has started. Previously
  it checked cancellation only before dispatching, so a cancelled send to an
  unreachable or stalled printer still blocked for the full timeout (10s by
  default). Now a queued send never connects, a pending connect is abandoned
  within about 50 ms, and a blocked write is cut off by shutting down the
  socket; all three throw CancellationError. A send whose bytes were all
  written before the cancel still reports success.
• ^SN serial numbers now appear in previews. The parser had no ^SN
  case, so a SerialNumber element (or any ^SN field, including a serialized
  barcode) silently vanished from the rendered image. A preview now shows the
  start value, which is what the first printed label carries.
• A field's ^A font no longer leaks into later fields. ^A applies to
  one field only; a field without its own ^A uses the ^CF default, or Font
  A at 9x5 dots when there is no ^CF (verified against Labelary). The renderer
  carried the last ^A font, size and rotation into every following field, so
  a field that relied on the default rendered in the previous field's font. The
  ^FW default rotation still applies.
• Font A rendered at about half its width. Its width was scaled by
  width / height like the scalable Font 0, but ^AAN,9,5 is Font A's native
  9x5 cell, not a squeeze. Each character now advances 6 * width / 5 dots,
  matching Labelary.
• Two data commands in one field produced two elements. A second ^FD or
  ^SN before ^FS now replaces the first, at the field's latest ^FO, as on
  Labelary. This was visible once ^SN rendered: a ^FX comment mentioning
  ^SN drew its own text at the top of the serial_number fixture.
• calibrateFull() was documented as a more thorough calibration. ~JG is
  Zebra's graphing calibration: the same measurement as ~JC plus a printed
  graph of the media sensor profile, so it is a diagnostic that costs extra
  labels.
• Tests: the loopback FakePrinter could close the same descriptor twice
  (once in shutdown(), again when its connection handler finished), which
  could close a socket belonging to another test running in parallel.
• Tests: refreshed 36 visual-test reference images that had drifted from
  current Labelary output. None of the drift was a renderer regression: ZPLKit
  agrees better with the fresh references in 27 of the 36. Ten graphic_*_fill
  references predated the switch from SF Symbols to FontAwesome icons, and
  seven fetched in 2026-07 were rendered from a percent-encoded request body,
  so every space in their label text came out as a literal +. The rest are
  small Labelary-side changes (a different QR matrix for the same data, and
  shape-edge rasterization). The four ^FR references, which are ZPLKit
  renders on purpose, were left alone. Visual accuracy rises from 90.9% to
  91.1%.
• Visual suite: features_textblock rises from 88.6% to 90.7% (the font
  leak). serial_number and retail_label change because they now draw
  their serial fields; nothing else moves by more than 0.1 point.

Changed

• PrinterError has a new case, printerBusy(host:port:). Exhaustive
  switch statements over PrinterError without a default need a new arm.
• CI: the WebAssembly job moved to the swift:6.4.1 container and the 6.4.1
  Wasm SDK. The Linux job already follows the floating swift:6.4 tag.