VED-011 marriage exact-reuse performance validation

Date: 10 October 2026. Status: accepted locally; focused, broader and both full real-kernel HTTP cases pass, with exact response equivalence.

This pass follows the native arithmetic optimization. Both passes remain local and uncommitted, on engine 03dc7e0 with the unchanged version 6.9.9 and required native binary. The earlier receipt retains the first pass's fingerprints and measurements as historical evidence.

Implementation and reuse boundary

The warmed four-day profile after native arithmetic optimization attributes 5.85 cumulative seconds to 2,794 repeated TT/TDB transformations, 2.66 seconds to 4,890 record-argument calculations, 2.07 seconds to 909 frame-angle calculations and 2.65 seconds to frame rotations. These are nested profiler times, not additive contributions or production latencies. The initial paired experiment showed that single-body cases benefit differently; a complete request is also measured before asserting practical improvement.

The bounded reuse standard defines the governing deterministic computation, source/owner boundaries and admission contract. One private, standard-library helper is independent of marriage. Five request-owned memos reuse clock conversion, record arguments, frame polynomials, complete frame angles and trigonometric pairs.

There is no rounded epoch key, interpolation, cross-request shared state, cached exception or cached mutable result. Exact key fields include the complete differential and coarse/precise mode where relevant. Storage is bounded and creation-thread ownership is enforced. Numerical/source policy, all existing tolerance and resource limits, and the public engine/REST contracts are unchanged. No native source or build change is made in this pass.

Source-bound admission and work counters remain outside the new pure memos. Their hit statistics measure removed arithmetic; they do not redefine public evaluation, reader or root-iteration counts. Code decomposition follows the existing Moira differential state and explicit frame polynomial/trigonometric objects; no secondary engine supplies a formula or execution pattern.

Validation plan and retained diagnostics

Receipt directory: C:\dev\outputs\ved011-marriage-reuse-2026-10-10. All execution uses the project Python 3.14.3 with downloads disabled, strict known-issue expiry and the existing discovered DE441/LE441 resource. The native binary and kernel remain those identified in the first performance receipt.

Focused checks cover exact adjacent-epoch/signed-zero keys, every differential field, zero/finite capacity and eviction, failure recovery, mutable payload rejection, separate dependency scopes, invalidation, cross-thread rejection, garbage-collection of finished scopes, and an actual non-marriage NAIF clock consumer. Marriage checks compare the original frame arithmetic schedule across the admitted date range and coarse/precise modes, compare all nine bodies and combined signals with reuse disabled, and retain source and work accounting. A warmed record-argument memo must not bypass a reader budget.

The existing rational derivative, real serving-record, frame, observer, source-seam and full-composition tests remain applicable. Both full HTTP cases ran again, including unavailable evidence at high latitude. Complete worked-case payloads were compared with the preceding accepted implementation; the native fingerprint is unchanged, and no response-field normalization was performed in this pass.

The first added non-marriage test used a nonexistent tt_to_tdb_jd name; it was corrected to the actual tt_to_tdb API. An initial cache-statistic assertion expected twelve polynomial hits where six are correct: the outer angle memo serves the repeated calls. Both are test-authoring repairs; no numerical tolerance, expected scientific result or existing fixture was changed. Initial failed receipts remain alongside subsequent acceptance evidence.

Complete baseline telemetry and paired timing

The unchanged first-pass implementation completed the full six-hour HTTP test again before this pass's acceptance. Its complete window and personal-direct responses exactly match the first performance receipt, without normalizing any response field. Telemetry retained an 8,192-entry observation window for each operation and counted exact repeated keys:

Operation	Calls	Repeated exact inputs	Time spent recomputing those inputs
TT/TDB differential	182,729	62,265	36.04 seconds
Record arguments	277,220	60,814	9.91 seconds
Frame angles	74,158	18,140	12.18 seconds

These measurements include the full test's additional interior validation; they are telemetry, not a claim that every repeated key survives the final cache capacities. The instrumented HTTP calculation took 553.27 seconds; the full test took 604.69 seconds. The earlier uninstrumented first-pass HTTP measurement remains 497.29 seconds. Concurrent focused validation and instrumentation prevent treating the newer baseline as a controlled latency comparison.

The reproducible paired benchmark runs the same warmed reader, fresh request scopes and alternating execution order, disabling only the new memos for its reference. It requires identical evidence, sources and public work counters before recording a timing. The final shared-sky workload uses 32 epochs and all nine bodies, both precision modes, plus combined solar/lunar signals. The first four-epoch diagnostic was too short for stable timing and is retained as an exploratory run, not the principal measurement.

$env:MOIRA_NO_DOWNLOAD='1'
$env:PYTHONUTF8='1'
.\.venv\Scripts\python.exe scripts/benchmark_marriage_reuse.py --output C:/dev/outputs/ved011-marriage-reuse-2026-10-10/paired-reuse-final.json --repetitions 3

paired-reuse-final.json contains all twelve successful pairs, exact evidence hashes, per-operation hits/misses/capacities and both execution orders. commands.json records the exact broader validation and full HTTP invocations.

Case	Median reference seconds	Median reuse seconds	Median paired speedup
Four-day Moon crossings	3.926	3.681	1.067x
Four-day Rahu crossings	0.403	0.366	1.088x
Four-day Venus crossings	1.094	0.929	1.171x
Shared sky, 32 epochs	2.365	1.796	1.326x

These are observed component timings; single-body search benefits less than multi-body composition. They do not predict one universal full-request gain.

Broader regression acceptance

The combined slice passes 372 tests, with 26/26 DE441 resource uses executed. It includes 29 new reusable-helper and marriage-admission checks, the existing complete marriage unit/server corpus, real geometry and budget checks, native parity/nutation/import, SPK/DAF, public-surface and docstring governance tests. All arithmetic and scientific thresholds remain unchanged.

Six unrelated optional checks skip: two legacy jplephem comparisons and four small-body file checks requiring centaurs.bsp, sb441-n373s.bsp or minor_bodies.bsp. No missing dependency or kernel was installed to run those unrelated checks. Both long real HTTP cases are selected separately.

Ruff and Python 3.10 grammar pass for all six computational/test/benchmark Python files in this pass. The final runtime/source state is fingerprinted before full HTTP acceptance. The benchmark workload was strengthened after the earlier exploratory measurements; numerical runtime code is unchanged.

Full HTTP acceptance

The six-hour Delhi window calculation completes in 409.195 seconds, versus 497.286 seconds at the preceding accepted native checkpoint: 1.215x faster, or 17.7 percent less elapsed time, from 8.3 to 6.8 minutes. The complete canonical JSON response matches exactly, with no fields removed or normalized. The personal direct request/response also matches exactly. The 45 cells, 46 bands, 62 evidence parts, 613 findings, 88 supporting certificates and all public work counters are preserved. This is regression equivalence under the same declared numerical model, not a new independent full-profile astronomical oracle.

The high-latitude HTTP assessment completes in 347.893 seconds, versus 442.404 seconds previously: 1.272x faster, or 21.4 percent less elapsed time, from 7.4 to 5.8 minutes. Its complete response also matches exactly. Unavailable Lagna and planetary apparition evidence, the original reasons, raw witnesses and not_certified status remain intact. The 167,429 evaluations, 1,113,819 reader calls and 59,056 root iterations are unchanged.

Both full tests pass in 821.33 seconds overall, with 4/4 DE441 resource uses executed and no skips. The individual tests take 459.00 and 357.93 seconds, including assertions beyond the timed HTTP calls. The personal direct worked case is byte-identical to its predecessor. full-response-equivalence.json records all three comparisons, raw receipt fingerprints and timings; its normalization lists are empty.

The validation runner reads cache statistics only when the ordinary request receipt is built. It does not instrument individual numerical evaluations. full-reuse-statistics.json retains these observations. Multiple receipt builds from the same request are repeated observations, not independent runs; the following table uses the final state of each request once:

Reused computation	Window-request hits	High-latitude-request hits
TT/TDB differential	60,789	51,359
Record arguments	60,814	53,874
Frame polynomials	21,293	19,963
Frame angles	14,282	9,852
Cosine/sine pairs	214,244	178,353

These are 371,422 and 313,401 successful memo hits, respectively. Each cache remains within its declared capacity, with 24,576 entries total across the five new memos at the end of each full request. Entry count does not imply a fixed process-memory byte bound. The original numerical and resource counters remain unchanged despite the eliminated helper computations.

These are observed whole-request timings on this host. The earlier baseline is historical and machine load is not controlled; brief documentation checks also overlapped the new window run. Paired same-process component measurements provide separate corroboration. Full synchronous calculations still take minutes; no interactive latency or service-level guarantee is implied.

Changed owners and lineage check

Owner	This pass's change
moira/_bounded_memo.py	Private bounded, thread-owned reuse with exact byte keys, immutable payload validation, failure retry and inspectable statistics.
moira/_muhurta_marriage_astronomy.py	Request-local TT/TDB and record-argument reuse; supplies the frame owner's cached trigonometric pairs to the original rotations.
moira/_muhurta_marriage_frames.py	Explicit differential key, shared precession polynomials, cached frame angles and cosine/sine pairs; preserves arithmetic order.
tests/unit/test_bounded_memo.py	Adversarial reusable-helper contracts and a non-marriage NAIF clock consumer.
tests/unit/test_muhurta_marriage_reuse.py	Original-schedule exact-bit comparisons, scope/reader lifetime, all-body composition, source isolation and budget-boundary checks.
scripts/benchmark_marriage_reuse.py	Reproducible alternating-order benchmark with exact evidence/source/counter admission.
Reuse standard and this receipt	Governing reuse contract, measurements, provenance, limitations and acceptance.
Marriage standard, prior performance receipt, implementation plan and remaining-work register	Link current implementation/acceptance while preserving historical checkpoints.

The implementation starts from the named immutable computational object and exact input identity. Keys contain every numerical state field rather than inferring equivalence from a sampled output. Frame decomposition follows the precession polynomials and trigonometric pairs already present in Moira's operation schedule. Reader admission, source boundaries, doctrine and event assembly stay in their existing owners. Inspection found no new external-engine formula, index convention, branch repair or secondary-parity authority claim.

The earlier native optimization, its tests and build settings are preserved. This pass adds no dependency, public selector, REST schema, version change, new VED-011 purpose, background service or website implementation. Source publication remains a separate action; these changes are local and uncommitted.

Final acceptance and fingerprints

The final latest-per-test reconciliation contains 374 distinct passes, six disclosed unrelated skips and no unresolved failures. The combined regression and full HTTP XML receipts own that count; repeated diagnostic and focused runs do not inflate it. Every selected real-kernel resource use ran.

The six computational/test/benchmark files and native binary match the source state frozen before full HTTP validation. Source fingerprints normalize UTF-8 text to LF; binary and evidence receipts use their actual byte fingerprints.

Receipt	SHA-256
source-state-final.json	d77fa50fa83ee1de7841ca60391f61be32857773fc77bb080cef69ae22fa4727
latest-results-final.json	d5d1540cef81c48a99a59256f1089b8d6e644f740c444fc8403df483be492735
full-response-equivalence.json	74631e00781a79bb187619a37e6124fc2fb14ebdc6d91f4a07fc5b08249dea1d
full-reuse-statistics.json	544fc68ace1eb6cb1d10a5c82e254404b57ce1424bb05206ce7fa1446ca892a9
paired-reuse-final.json	1099665ba9a537ed3bed41e170b62508d08bfe120e88fe3351ad53a6afbef97d

The original scientific fixtures and tolerances, required native binary and DE441/LE441 resource are unchanged. Final static gates cover scoped Ruff, Python 3.10 grammar, changed-line whitespace, release-facing documentation, release identity, Hellenistic and REST inventories, website publication metadata and canonical-to-generated wiki synchronization.