BowEcho v0.33.2

BowEcho v0.33.2 — "The Synthetic Radar Reference Release"

BowEcho v0.33.2 turns the WRF synthetic-radar workspace into a deeper,
inspectable forward-modeling system, while updating the embedded SimSat engine
to v0.2.0. It also fixes the P3/ISHMAEL edge cases and temporal-memory regression
reported during the v0.33.1 field test cycle.

Start here

• Open Windows > WRF, choose Real radar (balanced) — recommended, and
  build one or more raw wrfout times. The atmosphere is now Frozen by
  default, so ordinary builds do not require a neighboring scene or the heavy
  temporal cache.
• Choose Maximum fidelity (slow) when you deliberately want adjacent-time
  atmosphere interpolation during the scan. This changes the atmosphere sampled
  by later rays; it does not create extra loop frames.
• Choose P3/ISHMAEL T-matrix (research) only for native P3 mp_physics
  50–53 or ISHMAEL 55 output with the required property fields. It now starts
  Frozen as well. Temporal raw-state pre-closure remains an expert opt-in.
• Refresh current frame(s) reruns the remembered WRF sources using the
  controls currently shown. The first completed tilt appears as a preview, and
  long jobs can be cancelled without installing a partial final loop.
• Open Windows > SimSat for the embedded SimSat v0.2.0 renderer. Existing
  WRF, cached/automatic HRRR, Satellite-window, native-plot and PNG workflows
  remain available.

Scheme-native P3 and ISHMAEL radar

The research property T-matrix path no longer represents each native frozen
category with one characteristic particle. BowEcho reconstructs the supported
P3 or ISHMAEL particle-size distribution, integrates table-ready quadrature
nodes over that distribution, and reduces additive scattering and fall-speed
moments in a deterministic order.

The implementation preserves the source mass, number, rime, density, liquid
fraction and shape contracts; handles oblate/prolate and wet/dry branches;
keeps ZH, ZV, covariance, signed KDP, attenuation and fall-speed moments
additive through spatial sampling; and derives ratios only afterward. Exact
source/table applicability remains fail-closed. Unsupported physics, missing
properties, out-of-domain particles or a mismatched external table pack stop
the research run instead of silently substituting the bulk Rayleigh operator.

This release hardens the real WRF boundary conditions found during RC testing:

• tiny negative mass values caused by floating-point NetCDF output are cleaned
  only within the audited roundoff tolerance;
• native P3 qsmall, low-density, low-number and endpoint shapes are normalized
  or omitted through bounded, mass-audited rules;
• zero-ZH and subnormal fall-moment tails no longer produce impossible negative
  variance or nonzero moments without reflectivity;
• rain and mixed-phase closure uses the complete audited table domain and
  consistent material density; and
• a 0.1 K numerical tolerance handles temperatures immediately outside the
  standalone-rain LUT edge without turning into general extrapolation.

P3/ISHMAEL property output remains explicitly research-only and not independently
validated as an operational radar simulator.

Optional NVIDIA CUDA execution

The WRF radar controls now expose Auto, CPU, and NVIDIA CUDA compute
preferences. Supported native dry P3/ISHMAEL table evaluation can use a qualified
NVIDIA device; immutable science tables stay resident, requests are adaptively
coalesced, and buffers are reused across batches. CPU replay remains the exact
fallback for unsupported work or an unavailable device.

Fixture comparisons require exact output float bits between the CUDA path and
the CPU reference. Provenance reports the requested and effective backend,
device, batches, nodes, fallbacks and timing. CUDA does not make every job
faster: the CPU still owns native PSD reconstruction, ordered reduction and
fallback, and small volumes may remain dominated by submission overhead.

Radar validation and inspection

The WRF workspace adds tools for determining why a synthetic product looks the
way it does instead of treating the renderer as a black box:

• exact observed-scan replay preserves cuts, rays, timing, moment availability,
  Nyquist, PRT and missing sectors;
• linked Observed, Simulated and Difference panes compare overlapping gates;
• Algorithm Truth Lab scores supported algorithms against known synthetic truth;
• the gate inspector explains retained source cells, weights, instrument stages
  and unavailable evidence; and
• Ideal, Measured and Presented stage products can be emitted explicitly for
  controlled experiments.

Per-ray instrument metadata survives replay and export. Cancellation is checked
through inventory, scene reads, P3 preparation, cuts and CUDA work; cancelled
jobs do not replace the current completed radar loop.

Temporal memory and safer defaults

v0.33.1 introduced a separate 8 GiB temporal-build cap without migrating users
and allowed science presets to reset it. That made previously usable source sets
appear broken even when the normal playback-loop RAM setting was higher.

v0.33.2 fixes that regression:

• the temporal-build cap defaults to 64 GiB;
• the unmodified v0.33.1 8 GiB value migrates once to 64 GiB;
• deliberate custom values persist across updates and preset changes;
• the control is labeled Temporal build RAM cap in GiB and is explicitly
  distinguished from playback-loop RAM; and
• preflight failures identify the control and suggest fewer WRF frames or fewer
  retained diagnostic products.

The cap is a safety ceiling, not a RAM reservation or proof that the machine has
that much memory available. Instantaneous + Frozen avoids adjacent-scene
temporal processing. Real Radar and P3 Research now default Frozen; only Maximum
Fidelity enables adjacent-time interpolation automatically.

SimSat v0.2.0

BowEcho now embeds SimSat v0.2.0 at its public release revision. The reviewed
display baseline adds deterministic two-subcolumn cloud closure, a bounded 4.0
low-sun land-normalization gain, direct top-down camera-to-cloud atmosphere,
transmittance-space ground-shadow anti-aliasing, exposed-edge feathering and a
4096-step sun optical-depth ceiling.

The optional post-light terrain toe remains a labeled, default-off CPU display
experiment. NOAA heritage Natural longwave-IR grayscale is available without
silently replacing an existing persisted CIMSS selection. Compact production
caches remain SSB v6, so this update does not force another compact-cache rebuild.

Other fixes

• WoFS sounding draping uses stricter calibration and georeference handling
  instead of allowing a small successful subset to hide widespread failures.
• SPC mesoscale discussions expire against their actual validity rather than
  lingering as active products.
• WPC mesoscale precipitation discussions are ingested and presented alongside
  the existing discussion products.
• The in-app simulated-radar and SimSat guides describe the new controls,
  workflows, provenance and scientific boundaries.