moira

Physical Heliacal Visibility Phase 6 Closure

Date: 2026-08-04 Status: Complete; two bounded native kernels admitted; event budgets remain red Baseline implementation revision: 7e7e2886bc72dfd6e55231902ecd4a73d6b47832 Final measured implementation revision: 35c1f26b3fb5e44dd31bed168e1b30449ddc7a44 Evidence class: Numerical differential and performance evidence, not new scientific validation

Decision

Phase 6 admits exactly two doctrine-free dense numerical kernels:

1. physical_visibility_response_weights_v1, which evaluates the fixed piecewise-linear spectral color warp, normalized photopic and scotopic response weights, and their S/P ratio from Python-owned inputs; and
2. physical_visibility_direct_extinction_v1, which evaluates one already bracketed direct-extinction LUT row interpolation and its spectral transmission.

The event solver, event ownership, model identifiers, color-law selection, manifest and pack admission, validity-domain decisions, typed failure reasons, policy, provenance, receipts, and result construction remain in Python. The native bindings are private implementation details. The Python numerical implementations remain executable differential oracles.

No geometry cache, approximate profile binning, solver port, policy port, or public API expansion is admitted.

Why these kernels

The baseline and profiling receipts are:

• Phase 6 benchmark checkpoint;
• Phase 6 profile and census checkpoint;
• baseline benchmark artifact; and
• formal profile artifact.

The formal profile attributed only 0.78% of non-overlapping internal time to the event-solver implementation. Exact-input census runs found only 23 repeated horizontal inputs among 53,886 Jupiter calls and 12 among 37,423 Sirius calls. All 16,145 Jupiter target-profile contexts and direct-extinction altitudes were distinct. Those results rejected a solver port, exact-input geometry cache, and approximate profile reuse.

The profile did identify dense 400-bin spectral loops. Before native work, the microbenchmark measured about 277.8 microseconds per Jupiter dynamic profile resolution and 92.2 microseconds per direct-extinction interpolation. It also identified an unnecessary second 800-value validation walk after the immutable pack had already validated and generated the arrays. Phase 6 removed only that duplicate walk; the independently constructible TargetSpectralProfile continues to validate every supplied weight.

Exact differential admission

The admission harness is scripts/validate_physical_visibility_phase6_native.py. Its exact receipt is physical_visibility_phase6_native_validation_2026-08-04.json, SHA-256 5ce3eb7a5100c15e8322e625da17624489a0ef874895f466e4cd749441c8e11e.

The clean final revision used a freshly rebuilt native binary with SHA-256 ef5fedf3fc0c98b4343af02f63d966e9e3cc94f9f46805f57d1b98a18e7ddd4a, explicit DE441/LE441, and physical-visibility pack 1.2.0 with manifest SHA-256 cf93433a9f66a5ea92832271ce3c4b023fcc8693164803539a9f1be85b17468c. Normal execution remained offline under MOIRA_NO_DOWNLOAD=1.

The boundary-inclusive differential matrix contained:

• 14,205 target-response cases across Mercury, Venus, Mars, Jupiter, and Saturn;
• 1,001 phase points for each non-Saturn target;
• a 101 by 101 phase-angle and effective ring-latitude grid for Saturn;
• 2,001 target-altitude points across the complete 0.25 through 45 degree direct-extinction domain;
• 512 response and 256 direct cases repeated under eight-thread concurrency; and
• five 2,000-operation timing samples per Python/native kernel pair.

All checks passed. Observed maximum absolute differences were:

Quantity	Observed maximum	Admission tolerance
S/P ratio	8.881784197001252e-16	5e-15
Photopic response weight	1.734723475976807e-18	5e-18
Scotopic response weight	1.734723475976807e-18	5e-18
Response normalization residual	1.1102230246251565e-16	2e-15
Direct-extinction magnitude	0	2e-15
Direct transmission	0	2e-15

Serial and concurrent fingerprints were identical. The response-weight kernel was 6.945 times faster than its Python oracle at the median, and the direct- extinction kernel was 5.078 times faster. These are kernel-only measurements, not public service-level promises.

End-to-end performance gate

The final three-sample receipt is physical_visibility_phase6_native_admission_benchmark_2026-08-04.json, SHA-256 ec79280ebfd3653c81aaaf8eb566a12504539b5da9e0db34dbba6228213f28ad. It cryptographically binds the passing native-admission receipt, clean revision, rebuilt binary, and exact pack.

Workload	Baseline sizing sample	Final three-sample median	Budget	Result
First evaluated assessment	1.940 s	2.599 s	3.0 s	Pass
Warm assessment	0.269 s	0.325 s	0.5 s	Pass
Jupiter 30-day event	36.446 s	28.106 s	15.0 s	Fail
Sirius 20-day event	21.778 s	18.911 s	15.0 s	Fail

The baseline event values were one-sample sizing measurements, so the apparent 22.9% Jupiter and 13.2% Sirius reductions are directional rather than a formal cross-revision statistical estimate. A same-process diagnostic A/B also preserved the exact Jupiter result fingerprint while reducing the native path from 35.723 seconds to 29.289 seconds. The final gate deliberately exited nonzero because the two event budgets remain unmet.

The missed budgets are recorded performance limitations, not hidden failures. Phase 6 is an optional measured strengthening phase; its exit gate requires a truthful admission decision, complete differential evidence, and preservation of Python semantics, not arbitrary migration until a timing target turns green. Further large gains would require a separately justified attack on astronomical geometry or certified-search evaluation count. Neither is admitted by this phase.

Regression verification

The exact Phase 3 external-pack goldens passed at the final implementation:

2 passed in 55.1s
Jupiter exact event JD: 2460070.591375516
Sirius exact event JD: 2461255.950317484
Planetary resource identity: DE441 in all four acquisitions

A focused regression collection then passed 314 tests covering the pack and LUT, planetary and stellar targets, spectral composition, event behavior, public contracts, legacy policy, native import/runtime verification, physical REST routes, and general phenomena routes. It recorded 84 successful DE441 planetary-resource acquisitions, no skips, and no failures.

The final event benchmark retained the exact pre-Phase-6 result fingerprints:

• Jupiter: 2222fc44a89a23db2bf8f23fecdcd08e8bb07106b616c49ae6cf633463c8e863; and
• Sirius: 8f584f41084943ccaf1068a02c37561eeb469c7219b053709d3c41b8fdb9639e.

Exit gate

• Representative Python baselines and the benchmark environment are recorded.
• The native admission decision is recorded.
• Only profiled, doctrine-free numerical kernels moved to C++.
• Manifest, policy, domain, failure, provenance, event, and result semantics remain in Python.
• Python differential references remain executable.
• Boundary-inclusive admitted-domain differential grids pass.
• Invalid-domain behavior remains Python-owned and regression protected.
• Deterministic concurrency passes.
• Numerical tolerances and performance evidence remain separate.
• Authoritative event goldens and focused public/REST regressions pass.
• The still-red event budgets are explicit.

Phase 6 is closed. Phase 7 is the next implementation phase.

The implementation-plan status row and checkboxes require a narrow mechanical sync after the protected in-progress Phase 4/5 edits to that shared roadmap file are committed. This closure does not overwrite or absorb those unrelated working-tree edits.

This site is open source. Improve this page.