ephemerides-spectral

High-precision HDC reference instrument for the Sol Star System. Status: v0.26.0 — production-ready. Two-stage architecture: three interchangeable integer-ALU phase-residue encoders (bip Python / c native / complex128 Python reference) feeding an FPU complex64 HD pipeline (syzygy / observer-bind / eclipse-probability); 52-body roster (v0.16.0 Tier-1 expansion: +4 Saturnian Lagrange trojans (Telesto/Calypso at Tethys L4/L5; Helene/Polydeuces at Dione L4/L5 — first L4/L5 entries in BODIES, multiplicity-2 Laplacian degeneracy at host frequency) + 3 Jovian irregulars (Himalia/Pasiphae/Sinope) + Proteus/Nereid (Neptune sub-graph completion); resonance-graph multi-leg find_itn_chains (v0.17.0 — Dijkstra-style graph search over the (body, epoch) state space generalising the v0.8.1 closed-form Hohmann anchor to multi-leg pathways with small-integer (p, q) gear-ratio resonance signatures per leg; pure-Python addition built on top of find_itn_pathways, no ABI bump); bridge.body_architecture (v0.18.0 — first spectral-architecture surface, classifying heliocentric bodies into the canonical inner-8 / outer-5 partition via the resonance-weighted gateway-graph Laplacian Fiedler vector; the cyclic-group encoder discovers the asteroid-belt boundary without being told it exists; Pluto and Neptune share the deepest outer entry via their 2:3 mean-motion lock; pure-Python addition, no ABI bump; research origin notebook §13.8); bridge.predict_itn_accessibility (v0.18.1 → v0.18.2 — closed-form spectral Δv estimate from the §13.9 hybrid Fiedler-distance regression; v0.18.2 upgraded to a 2-D (f₂, f₃) embedding with R² 0.51 → 0.64 and LOOCV MAE 4.24 → 3.12 km/s; useful for fast first-pass triage at

ratchet-pinned; Rule 10 enforced via cross-platform pedantic-build CI matrix; Rules 6+7 manual audits clean); pre-merge docs+parity hygiene check (v0.13.3) as a soft-warning GitHub Actions workflow. See the Status section below for the version-by-version landing record.Overviewephemerides-spectral is a hyperdimensional-computing instrument that encodes the barycentric state of our star system using high-precision ephemeris data (NASA JPL DE441 / DE442) as resonant phases over a graph Laplacian.Two-stage architecture: integer-ALU phase residues, then FPU HD liftThe package separates the phase-residue computation (integer ALU, no FPU on the hot path) from the HD operations (syzygy projection, observer-bind, eclipse-probability — necessarily FPU because channel bases are unit-magnitude complex). The phase-residue stage has three interchangeable encoders; the HD stage runs complex64 natively or complex128 as the reference.Phase-residue encoders (integer ALU) bip (default) — bit-serialised integer ALU in pure Python. Phase composition lives in the cyclic group Z_{2^32}; binding is (φ_1 + φ_2) mod 2^32, which is implicit uint32 overflow — no FPU in the hot path. 305× faster than the FPU reference; 256 KB state at D=65536. Always available. c (v0.3.1+) — native C library (libephemerides_spectral.{so,dll,dylib}) bundled in the platform wheel under _native/, loaded via ctypes. Byte-for-byte identical phase residues to bip; ~1000× speedup on the chunk loop (encode at +20 yr: 46 ms Python → 0.04 ms C). Falls back transparently to bip if the binary isn't loadable (sdist installs without a C toolchain, Pyodide / WASM, the pure-Python fallback wheel).

All three phase-residue encoders implement the same algebraic substrate (cyclic-group representation of celestial phase-space, graph-Laplacian eigenbasis) and produce identical uint32[38] residues at any JD; they trade precision for speed (the integer-ALU encoders are exact in Z_{2^32}; the complex128 reference is float-ULP-quantised).HD pipeline (FPU, complex64 production / complex128 regression)Once the integer phase residues are computed, the HD operations (syzygy operator, observer-bind, eclipse-probability) lift them to a D-dimensional hypervector. This stage is necessarily FPU because the channel bases are (cos(φ), sin(φ)) complex pairs. backend="auto" (default for get_local_view / get_eclipse_probability) — selects c if the native binary is loaded (Tier 2b ABI v6), else falls back to bip integer phases lifted via the Python complex64 shim. backend="c" — Tier 2b in C (es_encode_state_hd / es_bind_observer / es_get_eclipse_probability); complex64 storage (single-precision); ABI v6 since v0.7.0; parity smoke pins both paths to within float-ULP. backend="fpu-ref" — Python complex128 reference encoder with unit-norm complex Gaussian bases. Regression baseline only; used to validate the c and bip paths against double-precision ground truth, not as the production HD path. Kept for the same reason scientific software keeps reference implementations alongside production ones. TL;DR on "pure ALU": phase residues are integer ALU end-to-end (BIP encoder hot path is uint64/int64/uint32, no floats); HD operations (syzygy / observer-bind / eclipse) lift those residues to complex64 and run on FPU. The package is not pure-ALU end-to-end — the HD pipeline can't be, because complex-magnitude bases require trigonometric channels. The integer-ALU discipline applies to the encoder hot path and is enforced by the JPL Power-of-Ten audit (Rule 10 pedantic-build matrix).

The moon set covers Earth's Moon, Mars's Phobos / Deimos, all 4 Galileans (Io, Europa, Ganymede, Callisto) plus the 4 inner regulars (Metis, Adrastea, Amalthea, Thebe), the canonical 9 Saturnians (Mimas, Enceladus, Tethys, Dione, Rhea, Titan, Hyperion, Iapetus, Phoebe) plus the Janus / Epimetheus co-orbitals, Uranus's Titania, and Neptune's Triton. Mean-motion resonances (v0.5.0+): 7 entries in RESONANCES — Jupiter–Saturn 5:2, Neptune–Pluto 3:2, Io–Europa 2:1, Europa–Ganymede 2:1, Mimas–Tethys 4:2 (Cassini Division), Enceladus–Dione 2:1 (powers Enceladus tidal heating), Titan–Hyperion 4:3 (Hyperion's chaotic rotation). Natural-resonance gear group: Z_60 = Z_4 × Z_3 × Z_5. Runtime kernel patching (v0.4.0+): Diagnosed-fiber overlay — patches sit beside the published kernel as DATA, not code edits, and contribute per-body residue deltas at encode time. Inspired by Linux ksplice / kpatch; the kernel's published bytes never change. Bridge surface: apply_patch(name) / apply_custom_patch(...) / clear_patches(). Three patches in the bundled CATALOG authored from the v0.3.1 FFT residual analysis. v0.5.1 patch-shrinks-residual benchmark measured the catalog and showed partial vindication: J–S coupled patch shrinks both bodies' residuals by ~77% with phase-recovered authoring (research-side; stays out of the v0.5.x catalog until ≥80% on every body); Mars stays stuck at 3% due to FFT bin leakage. v0.5.2 adds windowed FFT + multi-bin patches for full predictive power. SPICE-free runtime (v0.5.0+): pip install works out of the box — both backends use codegen-baked initial phases shipped in _data/initial_phases.json.