v0.2.0

Added

• Interactive JTAG debug server (--jtag-server) (#124). jacquard cosim --jtag-server <PORT> opens a live remote_bitbang JTAG socket alongside the
  running GPU co-simulation, so an external debugger (OpenOCD, then gdb on top)
  can attach and drive the design's RISC-V Debug Module — halt/resume/step,
  read/write GPRs/CSRs/PC/memory, and load firmware — in lock-step with each
  debug transaction. The interactive sibling of --jtag-replay (recorded
  streams). Adds --jtag-reconnect (re-accept a debugger without restarting
  cosim) and a jacquard jtag-openocd-config helper that emits the matching
  openocd.cfg. Host-side only (no kernel changes); the CI gate is a real
  OpenOCD attach. See docs/jtag-debug.md.
• reg_init register value-injection for cosim (#108). A new
  reg_init array in the testbench JSON deposits a definite value into
  chosen registers at tick 0 with $deposit semantics — the seed clears
  the power-up X-mask, then the design's own logic drives the register
  normally (NOT force, which would pin a CDC crossing register and write
  zeros across the handshake). Each entry is { "name", "value", "width" };
  a multi-bit register resolves name[0]..name[width-1] independently.
  This is the register sibling of sram_init and the fix path for
  X-poisoned unreset CDC launch registers (#102): depositing on the launch
  flops lets X-aware cosim of debug-loaded firmware proceed. Composes with
  the x-assert detection work (#106). Regression: tests/xprop_cosim/
reg-init mode (cosim A/B — deposit clears the X that xprop mode
  requires to persist).

Fixed

• cosim on SRAM-less designs no longer panics. A design with zero SRAM
  produced a nil MTLBuffer (new_buffer(0)) whose .contents() is null;
  the SRAM data buffer is now sized to max(1) word (matching the X-mask
  shadow buffer), so pure-logic designs run under jacquard cosim.
• bi_24t bidirectional pad core read-back (#96). The AIG modelled a
  bidir pad's core read as Y = PAD always, so a design that drives a pad
  and reads it back in the same path read the external PAD net instead of
  its own drive A — wrong in two-state (reads the external stim/0) and
  conservatively-X under --xprop (the external pad is an undriven input).
  AIG::from_netlistdb now builds the tristate read-back combinationally as
  Y = OE ? A : external: on OE=1 the core reads its own drive A
  (X-exact — known whenever A is), on OE=0 it reads the external net.
  in_c/in_s input pads are unchanged. See ADR 0016. Unit test:
  aig::gf180mcu_chip_top_tests::bi_24t_models_tristate_readback.
• Empty-partition flatten panic (#128). A partition with zero boomerang
  stages — which mt-kahypar can emit at high partition counts (finer
  partitioning or a large --num-blocks) — panicked in
  FlattenedScriptV1::from ("index out of bounds: len is 0"). Such partitions
  do no work and are now dropped before flattening.