v0.3.0-rc.3

Added

• Behavioral-RTL on-ramp (ADR 0021). jacquard sim and jacquard cosim
  now accept behavioral Verilog / SystemVerilog directly — synthesis is
  transparent and cached, run through an embedded YoWASP Yosys (with the
  yosys-slang SystemVerilog frontend) via Rust + wasmtime. Built into
  release binaries via the synth feature; --yosys-wasm <PATH> overrides
  the bundled wasm. See docs/accepted-rtl.md.
• Cell-model IR descriptors (ADR 0019). Standard-cell libraries are now
  consumed as generated JSON descriptors — pin directions, combinational AIG,
  sequential roles, and timing — produced from Liberty by the
  liberty-to-cellir converter and embedded at build time. Adds SKY130
  (via a .lib.json reader) and IHP SG13G2 as a new built-in PDK with zero
  per-PDK Rust; proprietary libraries simulate via --cell-descriptor
  (no Jacquard build). The runtime binary is now self-contained for standard
  cells — no vendored-PDK read at simulation time — and the PdkVariant
  enum, per-PDK stdcell classifiers, and build.rs pin-table generation are
  retired. See docs/adding-a-pdk.md.
• Plural QSPI memory + writable PSRAM (ADR 0013). The SPI-flash peripheral
  went plural: qspi_memory: Vec<QspiMemoryConfig> with N independent
  instances (the legacy flash key folds into instance 0), N-instance GPU
  kernels on Metal + CUDA + HIP with independent backing stores, and an
  opt-in writable QSPI-PSRAM (RAM) mode (APS6404L-class: enter-QPI /
  quad-write / quad-read). Enables post-PnR cosim of chips whose main RAM is
  external QSPI PSRAM. Unset options ⇒ byte-identical to the read-only flash.
• GPU frame capture (Metal). JACQUARD_GPU_CAPTURE=<path> (with
  METAL_CAPTURE_ENABLED=1) brackets an MTLCaptureManager scope around a
  bounded window of cosim batches and writes an Xcode .gputrace for
  per-dispatch GPU analysis. JACQUARD_GPU_CAPTURE_SKIP / _BATCHES select
  the window. See docs/gpu-capture.md. (#174)
• Cosim perf report. jacquard cosim reports a per-edge CPU/GPU timing
  breakdown including ground-truth GPU-execution time from device
  timestamps (Metal GPUStartTime/GPUEndTime) — free of the distortion a
  full GPU trace imposes on thousands of tiny dispatches per batch.
  --cosim-perf-json <PATH> emits it as JSON for CI. See
  docs/cosim-perf-report.md. (#175)
• RTL-source provenance. sverilogparse now captures (* src *)
  attributes and carries them through the netlist and AIG; jacquard xsources
  reports the RTL source location of each design X-source.
• TNS / THS timing metrics in --timing-summary and --timing-report.
  Alongside the existing worst-single-slack WNS/WHS, the report now carries
  Total Negative Slack and Total Hold Slack — the sum of every negative
  setup / hold slack across the run (stats.total_setup_slack_ps /
  stats.total_hold_slack_ps). Additive JSON-schema change, bumped to
  1.2.0 (older reports still parse via #[serde(default)]). Salvaged from
  the timing logic in #17. (#9)
• JACQUARD_VENDOR_DIR environment override for the vendored-PDK root.

Changed

• jacquard build folded into sim/cosim. The standalone build
  subcommand is removed; behavioral RTL is synthesized transparently by
  sim/cosim (see the on-ramp entry above).