v0.2.0 — Multi-architecture modernization (Linux x86-64)

What's new in v0.2.0

All platforms, all modern CPU instruction sets

The original application was built for a ~2008 baseline (SSE2, no FMA,
old compiler). Wisteria ships one build per platform and microarchitecture
so every host runs the newest instructions its CPU supports:

Package	Contents	Requires
x86_64-linux tarball	baseline + AVX2 + AVX-512	any 64-bit Intel/AMD
aarch64-linux tarball	ARMv8-A (NEON)	Jetson, RPi 5, Ampere
windows-x64 zip	self-contained Julia port exe	Windows 10/11

Measured on the validation host (AVX2): the semicoherent search stage runs
2.1x faster than baseline (10.5 s vs 22.3 s) — and unboundedly faster
than the original no-SIMD application, which never used vector units at
all outside its FFT library. The Windows build JIT-compiles to the host
CPU on first run (AVX2/AVX-512 automatically).

Full pipeline, not just kernels

Photon input, barycentering with the JPL DE430 ephemeris (Roemer +
Einstein + Shapiro delays), semicoherent sliding-window FFT search over
(f0, f1dot) with interbinning, coherent follow-up, toplist bookkeeping,
and BOINC checkpointing (per sky position + on request; kill -9 safe).

Numerical fidelity

Cross-validated against the upstream application on identical input:
FFT spectra agree to float32 precision (1.1e-7 over 4.19M bins), toplists
identical (50/50). Every build passes the same injected-pulsar recovery
test on all three platforms (x86-64, aarch64 under qemu, Windows exe).

Packaging

Each archive contains the binaries + app_info.xml + app_config.xml +
INSTALL.txt. Linux builds are statically linked (FFTW, BOINC API, C++
runtime). The Windows exe is self-contained (Julia runtime embedded,
~93 MB zip, starts in ~1 s). Fill in your host's plan_class in
app_info.xml — see INSTALL.txt.