Native Fixed-Star Heliacal Parity Validation

Date: 2026-07-12

Boundary hardening: 2026-07-14

Status: admitted native accelerator; Python remains the governing manuscript

Governing Product

The product is the first observable morning appearance or last observable morning appearance of a sovereign fixed star under the declared Arcus Visionis and heliacal search policy. Python owns identity resolution, policy, public vessels, metadata, and fallback semantics. Native code accelerates the repeated astronomical observables and bounded daily search.

Defect Repaired

The former native visibility path called a five-term routine labeled as truncated IAU 2000B. It was neither the complete standard model nor equivalent to Moira's authoritative Python path. Native setting also omitted setting_elongation_threshold and setting_visibility_factor, rising metadata reported the setting threshold, signed elongation omitted the Sun's annual aberration/frame bias, fixed-star evaluators introduced a different geocentric parallax convention, and one Delta-T value was held over the search.

The replacement:

1. Loads the packaged IERS Table 5.3a/5.3b 2000_R06 coefficients in Python.
2. Publishes one immutable native series snapshot with radians-native results.
3. Uses the shared series in both the public binding and visibility engine.
4. Applies the same fixed-star propagation and apparent solar correction semantics used by the Python elongation and altitude products.
5. Carries setting threshold/factor doctrine, rising metadata, and a bounded linear Delta-T drift through native search.
6. Keeps coefficient loading lazy and makes raw uninitialized native calls fail clearly.

No second coefficient table or new runtime dependency was introduced.

Boundary Hardening

The admitted accelerator now enforces two additional invariants shared with the Python manuscript:

1. A forward search never returns or remembers a twilight event earlier than jd_start. This applies to rising, setting, single-star, and catalog-batch searches while preserving the existing civil-day day_offset meaning.
2. A native SPK evaluator is bounded by the inclusive coverage in its own DAF descriptor. Exact descriptor endpoints remain valid; out-of-range requests raise instead of clamping to the first or last Chebyshev record or allowing Hermite extrapolation.

Native heliacal dispatch additionally requires one descriptor to cover the complete TT interval evaluated by the daily search. If that admission test fails, Python resumes epoch-by-epoch segment selection and raises on a real coverage gap. The reproduced DE441-end setting search now fails explicitly in both Python and native-enabled policy modes rather than returning a synthetic post-coverage event.

HeliacalSearchPolicy now rejects non-finite or non-positive setting thresholds and setting visibility factors outside (0, 1]. The historical visibility_tolerance field has no source-owned computational meaning; it is therefore retained only as a compatibility field fixed at 1.0, rather than silently accepting values that do not change the product.

Differential Evidence

The independent event corpus uses Python with use_native_heliacal=False and native with use_native_heliacal=True; it is not native-versus-native testing.

• Stars: Sirius, Regulus, Capella, Acrux, and Arcturus.
• Event kinds: heliacal rising and heliacal setting.
• Latitudes: -30, 0, 31.2, and 60 degrees; longitude 29.9 degrees east.
• Epoch/search: 2024-01-01, 400-day forward window.
• Event-set requirement: exact found, not-found, magnitude-skip, and latitude-skip equality.
• Event time tolerance: 0.05 seconds.
• Signed elongation tolerance: 0.00001 degree.
• Metadata requirements: exact day offset and doctrine-owned rising/setting threshold.
• Custom setting case: 20-degree eligibility threshold and 0.8 disappearance factor.

The named Sirius observable probe at Alexandria additionally measured native versus Python residuals below 0.008 arcsecond in signed elongation, below 0.001 arcsecond in stellar altitude, and below 0.005 second in the solar twilight root for the sampled 2024 epochs.

The core series is compared with the scalar Python table evaluator to 1e-13 degree and with ERFA nut06a under the existing 0.001-arcsecond integration tolerance. Cold initialization, invalid replacement, concurrent readers, native argument rejection, single-star dispatch, adversarial native runtime, and server packet paths are also exercised.

On 2026-07-14 the ERFA corpus's historical anchors were corrected to proleptic-Gregorian 500 BCE (astronomical year -499, JD 1538803.5) and 200 BCE (astronomical year -199, JD 1648376.5). ERFA cal2jd, Moira julian_day, and Moira's inverse calendar conversion now jointly guard those identities. The corrected 12-epoch integration slice passed all 106 tests at the unchanged 0.001-arcsecond threshold; the worst corrected precession- nutation matrix residual was 0.000938 arcsecond.

The 2026-07-14 boundary corpus also exercises raw type-2 and type-13 evaluators at both exact descriptor endpoints and one second outside them, the guarded Python/native type-13 differential path, forward-start probes immediately after known rising and setting events, and the DE441 terminal-coverage reproduction. Its consolidated focused gate passed 153 tests; two optional jplephem comparator tests skipped because that development-only comparator was not installed.

Performance Evidence

Local Windows project-.venv measurements after kernel warm-up:

Product	Python oracle	Native default
Five-star setting batch, 400 days	5.31 s	0.97 s
Magnitude-3 setting batch, 175 searched stars	>120 s (timed out)	20.69 s

The native batch owns an explicit native_heliacal_workers policy. Its default is eight workers, bounded to 1-64, and results retain deterministic input/event ordering. Repeated noon nutation values live only in a request-scoped cache; there is no process-global astronomical-result cache.

These timings are performance evidence, not astronomical validation.

Verification Receipt

The final focused native/R06/ERFA/star/packet slice passed 170 tests under strict known-issue expiry. A separate relevant visibility/server slice passed 140 tests. The broader historical visibility integration file was also run and reported out-of-scope planet/Moon authority-corpus failures in Babylonian Venus and Yallop lunar cases. Those branches do not dispatch through the fixed-star native search; they remain an explicit repository validation gap and are not counted as passing evidence for this work.

Remaining Boundary

This admission covers sovereign fixed-star heliacal rising and setting only. Planetary, lunar-crescent, acronychal, and other generalized visibility products retain their existing Python-governed implementations and validation status. Native availability never changes public result-vessel meaning.