VED-011 marriage performance validation

Date: 10 October 2026. Status: optimization accepted locally; complete native, marriage and full HTTP acceptance passed.

This receipt preserves the first native-arithmetic performance checkpoint. The subsequent exact-reuse pass records further changes and their separate acceptance; the measurements and fingerprints below remain historical to this first pass.

This follows the source-published marriage package at engine 03dc7e0e1f5f2668cb832de11c8e2c72250d2d4d, generated wiki 2ce8edc1e9d6c5c07f1e94fc7faf626d2743f123. The original validation receipt retains its original measurements, failures and source fingerprints. This work changes execution cost, not the selected classical rules or the meaning of astronomical certification.

1. Measured bottleneck and implementation boundary

A warmed DE441 four-day Moon/Rahu/Venus crossing profile attributed 69.77 of 92.28 cumulative seconds to the interval-Chebyshev recurrence: approximately 76 percent of this profiled workload. It made 15,534 recurrence calls and hundreds of millions of Python calls, largely for immutable interval and differential operations. These profiler timings are diagnostics, not a production latency benchmark.

The native helper translates the existing _cheb_python operation schedule. Python retains the reference, serving record selection, time/coordinate policy, light-time branch families, interval isolation, source findings, exception targets, completeness decisions and serialization. Native code fuses only the repeated arithmetic for one coefficient series and its derivative. Its input/output fields remain the named differential's value, rate, common center, radius and centered flag.

Both rounded recurrence orders, a+b-c and a+(b-c), remain in the hull. Every outward rounding, whole-domain primitive-error guard and centered intersection is retained. The native translation unit disables contraction, reassociation and link-time rewriting locally; unrelated native kernels retain their build settings. The helper refuses non-nearest host rounding, non-gradual underflow and invalid numeric states. It retains the existing conditional arithmetic model; it does not establish platform-independent formal verification or physical ephemeris accuracy.

There are no changes to tolerances, time steps, work limits, counter accounting, supporting-history domains, public Python/REST selectors or classical profile coverage. No new dependency or data resource is introduced.

2. Reproducible paired benchmark

$env:MOIRA_NO_DOWNLOAD='1'
$env:PYTHONUTF8='1'
.\.venv\Scripts\python.exe scripts/benchmark_marriage_enclosures.py --output C:/dev/outputs/ved011-marriage-performance-2026-10-10/paired-crossing-benchmark.json --repetitions 2

The standalone benchmark uses the same warmed serving reader with fresh request caches, alternating Python/native execution order. All roots, uncertainty intervals, gap records, supporting certificates and work counters must serialize identically before a timing pair is recorded. It temporarily selects the Python reference only inside its own single-threaded process.

Four-day case	Python reference, seconds	Native helper, seconds	Speedup
Moon, first pair	17.349	4.587	3.782x
Moon, reversed pair	17.826	4.604	3.872x
Rahu, first pair	1.214	0.460	2.641x
Rahu, reversed pair	1.239	0.451	2.748x
Venus, first pair	3.376	1.208	2.794x
Venus, reversed pair	3.576	1.250	2.861x

These are component performance measurements and exact regression equivalence, not an independent full-profile oracle or a universal speed guarantee. Raw results and diagnostic profiles are retained under C:\dev\outputs\ved011-marriage-performance-2026-10-10.

3. Numerical and native-boundary evidence

The project Python 3.14.3 rebuilt the required native extension using python -m pip install -e . --no-deps --no-build-isolation; the full command uses .venv\Scripts\python.exe. Version remains 6.9.9.

The initial native-primitive slice passed 23 tests with three executed DE441 resource uses and no skips. It covers:

• Exact binary-state parity, including signed zero, on centered and uncentered arguments, four radii, record endpoints, 0–32 degree coefficient series, small/large magnitudes, subnormal inputs and cancellation.
• Independent rational values and first/second derivatives of expanded T1, T2, T4 and T8 polynomials at 33 rational arguments each; bounds must contain exact values.
• Eight actual DE441 routes at four epochs spanning the admitted 1900–2100 domain, three endpoint/interior argument boxes and every axis, compared with the unchanged Python recurrence without any tolerance.
• Rejection of empty/nonfinite coefficients and invalid interval/radius states, plus preservation of explicit numerical-unavailability errors.

The optimized backend fingerprint is 12b5552f993954a46e3e7c2a1e90af767ac312c3cbf9ba5765ad9cb85e1f55bb. The serving DE441/LE441 kernel remains 9a824e6a8ce3d39f8a513b17278ff7b7ad226f964c76d2019b0a51d8808d6ff5.

An isolated native executable also compiles the exact helper source with MSVC 19.50.35730.0 and the same strict floating-point options. Nearest rounding succeeds; downward, upward and toward-zero rounding are rejected. Separately enabling x86 flush-to-zero and denormals-are-zero is rejected; restoring the original model succeeds. The seven host-mode checks are in arithmetic-host-check.log, with the standalone harness sources alongside it. They exercise real host modes, not substituted Python exceptions, and run outside the server/test processes.

The combined marriage, public-surface, native parity, nutation, SPK and DAF regression slice passes 343 tests, with 24/24 DE441 resource uses executed. Six unrelated checks skip: two optional jplephem comparisons and four legacy small-body file checks (centaurs.bsp, sb441-n373s.bsp, minor_bodies.bsp). No dependency or kernel was installed for those checks. The two full HTTP tests are deliberately selected separately. The XML and log are retained in combined-regression.xml and combined-regression.log; commands.json records the exact build, benchmark and test invocations, selected files, interpreter, working directory and environment.

4. Full composition acceptance

The original six-hour Delhi HTTP window calculation completes in 497.29 seconds, versus 1,701.96 seconds in the pre-change receipt: an observed 3.42x speedup, from 28.4 to 8.3 minutes. Its complete response is identical after changing only search_receipt.native_backend_sha256 to the previous identity for comparison. It retains all 45 cells, 46 bands, 62 evidence parts, 613 findings and 88 supporting certificates. The 188,201 evaluations, 1,286,318 reader calls and 65,877 root iterations are unchanged; the widest band remains 3.745664656 seconds. The result remains a completed restricted search with no eligible or indeterminate cells.

The high-latitude HTTP instant completes in 442.40 seconds, versus 1,510.44 seconds previously: 3.41x faster, from 25.2 to 7.4 minutes. Its complete response also matches after substituting only the native backend fingerprint. Lagna and both planetary availability contexts retain their unavailable states, with the exact original Jupiter/Venus undefined-horizon reasons and all raw witnesses. It retains 167,429 evaluations, 1,113,819 reader calls and 59,056 root iterations; incomplete coverage remains explicitly not_certified.

Both full HTTP tests pass: the complete tests took 565.53 and 457.31 seconds respectively, including their assertions beyond the measured HTTP call. All four selected DE441 resource uses executed, with no skips. The personal direct request/response worked-case file is byte-identical to its original. full-response-equivalence.json retains the exact comparisons and hashes for all three worked cases. Elapsed timing lives outside the calculation responses. End-to-end timings compare the archived run with this validation run, whose machine load is not controlled. The paired component benchmark above provides same-process, alternating-order corroboration of the speed gain.

Full calculations still take minutes on this host; this is a measured optimization of the complete synchronous calculation, not a background-job service or a latency guarantee for every admitted request.

5. Changed owners and scope

Owner	Change
moira/_muhurta_marriage_astronomy.py	Retains the Python recurrence and delegates its exact operation schedule through the private native bridge.
src/native/include/marriage_enclosure.hpp, src/native/src/marriage_enclosure.cpp	Defines the named differential-state transport and executes the fused outward-rounded recurrence with host-model guards.
src/native/bindings/moira_native.cpp	Binds the private primitive; curated public engine/REST contracts are unchanged.
CMakeLists.txt	Adds the numerical translation unit with locally strict floating-point and no-LTO settings.
tests/unit/test_muhurta_marriage_enclosure.py	Adds bit-exact Python-reference parity, independent rational derivatives, actual serving-record and invalid-state checks.
scripts/benchmark_marriage_enclosures.py	Provides the reproducible paired benchmark with complete-result/counter equality as an admission condition.
Source standard, this receipt and remaining-work register	Document ownership, measurements, acceptance and operational limits.

The governing object is the existing named differential state. The native schedule follows its explicit Moira reference and keeps both rounding-order branches; its compact transport follows those declared state fields. Source policy and astronomical assembly stay in their existing Python owners. Independent rational arithmetic and serving-coefficient evidence support correctness; timing and archived-payload equality have their separate performance/regression roles. No secondary astrology engine supplies this implementation's formulas or thresholds.

This checkpoint is local implementation work. No commit, push, release, deployment, website adoption or additional VED-011 purpose is included.

6. Final acceptance and fingerprints

The latest-per-test reconciliation has 345 distinct passes, six disclosed unrelated skips and no unresolved failures. Repeated primitive and rational cases count once. The four strengthened second-derivative cases also pass in rational-second-derivative.xml. The computational source and native backend stayed fixed throughout full HTTP acceptance; the stronger test assertions do not change that executable path.

Receipt	SHA-256
source-state-final.json	bb30f3b0d5808061f6dbae326248a201b96651f33c1865e8649300b5d8342c02
latest-results-final.json	1405f0316b9af388b9bbf889fee7b99d8a3844692ea0d3b0aea949106e12a0f9
full-response-equivalence.json	617d926871ec130ac9c1d1ec8ccd174b3beb70bdae744f44cfb5963ace23b699

The source manifest fingerprints all seven changed runtime/build/test/benchmark files using normalized UTF-8/LF text, and separately identifies the serving native binary. Scoped Ruff, Python 3.10 grammar, changed-line whitespace, documentation consistency, release identity, Hellenistic generated inventories, REST inventory, website publication metadata and generated-wiki synchronization all pass as the final static/documentation gates. Original scientific baselines and their tolerances remain unchanged.