Date: 12 October 2026. Status: All seven selected stages complete; authorized source-publication checkpoint below. Baseline: engine main, f8e9a615f04659c77a03713b302cb7189f4d890c, version 6.9.9; project .venv, Python 3.14.3, Windows. Scope: executed plan, source packet, public standard.

No native algorithm or dependency changed. Direct evaluation requires no ephemeris, resource reader, birth-chart construction or network. Normal REST application startup retains its existing engine lifecycle; that separate resource cost is not attributed to the sign/degree arithmetic.

The fixture directory owns small numerical transcriptions, not redistributed books. Its source manifest identifies local-library files and primary author URLs. The local corpus is C:\dev\ASTROLOGY-BOOKS-DATABASE. Source identity includes:

Every executable rule has its own source coordinates in discovery. Research-only BPHS/PVR/Narayana witnesses retain their separate identities and limitations. Hashes prove witness identity, not doctrinal correctness.

Validation distinguishes these evidence classes:

• All 24 printed method-2 sequence rows / 288 cells are literal independent arrays covering every Lagna, parity and modality. Three method-1 and seven method-2 duration examples are source checks.
• The twelve-value female recipe fixture stores both the printed Cancer ten and the selected own-sign-prose twelve. Eleven printed durations agree; the twelfth is a disclosed source conflict, not a twelve-of-twelve published oracle.
• The BPHS Argala chart uses the publisher's erratum adding Rahu in Gemini. The uncorrected omission is retained as evidence. It does not resolve the separate BPHS Chara diagram/table conflicts in the other volume.
• PVR Exercise 16 provides the full twelve-reference/eight-position occupancy table. Rath provides additional pair/factor witnesses and source-stated judgments, not a complete ranked strength algorithm.
• Independent exact-rational arithmetic checks degree-quarter boundaries and cumulative fixed-day projection. Synthetic longitudes never become claimed historical source degrees. All sixteen directed quarter combinations and 144 boundary cases (twelve signs, four offsets, three adjacent floats) are exercised, with mixed-matrix and special-third counterexamples.
• Twenty captured pre-repair Argala results and both legacy Chara cycle counts establish compatibility only. Source judgment, regression and transport parity are different claims.

C12 fixture correction: The preliminary Rath fixture asserted a negative Jupiter aspect to Gemini, an assertion absent from the article. Independent Jaimini geometry puts Pisces, Gemini and Sagittarius in the dual group: Jupiter in Pisces aspects both compared signs. The fixture now records both as true, retains the article's stated winner separately, and expects an indeterminate runtime comparison. No asymmetric aspect or arbitrary factor weight was added.

Two synthetic count tests initially put Ketu in the reference sign and thereby activated the selected reversal. They were corrected to isolate count comparison; separate all-sign/node/house-reference tests verify reversal itself. Neither correction regenerated expected answers from production output.

The final boundary review also established exact binary-input Moon-phase ownership near conjunction/opposition. Fractional arithmetic prevents subtraction from rounding a just-bright-half input onto the dark-half boundary. Character tests cover this alongside node-associated Mercury and sign-only uncertainty.

All commands use the owning .venv with downloads and external test networking disabled. tests/KNOWN_ISSUES.yml is empty and strict expiry is enabled.

$env:MOIRA_NO_DOWNLOAD = '1'
$env:MOIRA_TEST_MODE = '1'
$env:MOIRA_STRICT_KNOWN_ISSUES = '1'
.\.venv\Scripts\python.exe -m pytest tests/unit/test_jaimini_extended.py tests/unit/test_vedic_chara_d60_admission.py tests/server/test_vedic_chara_d60_admission.py tests/server/test_vedic_direct_input_hardening.py -m 'not external_network' -q

.\.venv\Scripts\python.exe -m pytest tests/unit/test_chara_profiles.py tests/unit/test_argala_assessment.py tests/unit/test_jaimini_profile_admission.py tests/server/test_server_jaimini_profiles.py tests/unit/test_jaimini_extended.py tests/unit/test_jaimini.py tests/unit/test_vedic_chara_d60_admission.py tests/server/test_vedic_chara_d60_admission.py tests/server/test_vedic_direct_input_hardening.py tests/server/test_server_jaimini_service.py tests/server/test_server_jaimini_routes.py tests/unit/test_api_surface_adversarial_audit.py -m 'not external_network' -q --junitxml=C:\dev\outputs\ved013-implementation-2026-10-12\targeted.xml

.\.venv\Scripts\python.exe -m pytest tests/unit/test_chara_profiles.py tests/unit/test_argala_assessment.py tests/unit/test_jaimini_profile_admission.py tests/server/test_server_jaimini_profiles.py -q --junitxml=C:\dev\outputs\ved013-implementation-2026-10-12\final-focused.xml

.\.venv\Scripts\python.exe -m pytest tests/server/test_server_jaimini_profiles.py -q -k 'not real_application_lifecycle' --junitxml=C:\dev\outputs\ved013-implementation-2026-10-12\final-transport.xml

.\.venv\Scripts\python.exe -m pytest tests/unit/test_jaimini_profile_admission.py tests/unit/test_api_surface_adversarial_audit.py -q --junitxml=C:\dev\outputs\ved013-implementation-2026-10-12\final-exports.xml

Deduplicating JUnit (classname, name) identities across the last four runs gives 847 distinct passing tests. The runs overlap; their totals must not be added as distinct coverage. This is a targeted package/shared-owner receipt, not a full repository or all-Vedic suite run.

Source fixtures cross owning modules, moira, moira.facade, moira.vedic, Moira methods, thin services and real in-process HTTP. Responses are compared including evidence, reasons and receipts. Hostile cases cover boolean/string/ nonfinite/huge numbers, missing and extra bodies, wrong precision, unknown IDs, unsupported switches, duplicate/excess references, projection overflow and forged input/request/result objects. Six-profile concurrent execution remains deterministic. The direct path is guarded against resource/network evaluation.

The real application lifecycle test explicitly restores the unpatched engine factory, checks the created Moira object, and exercises discovery and both new POST routes. The combined run recorded 29 executed resource checks using the available DE441 substrate; none of its tests skipped. Direct arithmetic tests and the benchmark do not require that kernel. The final transport subset excludes the already-passed lifecycle test by selection, not by a skip or degraded result.

Malformed engine result types, failed result invariants and invalid response contracts are internal failures. Injected defects return HTTP 500; they are not misclassified by the application's generic ValueError-to-422 handler. Actual invalid inputs retain explicit engine rejection and HTTP 422.

The benchmark tool records first-call and thirty warm repetitions, full response sizes, pair counts and implementation hashes in the performance receipt. Every iteration preserves exact engine/HTTP output equality. It uses a deliberately reader-free application factory; startup cost is covered separately above.

.\.venv\Scripts\python.exe -m tests.tools.benchmark_jaimini_profiles --output tests/artifacts/benchmarks/jaimini_profiles_2026-10-12/receipt.json

The fixture uses the maximum 22 references, with a populated 4-by-5 contested occupancy split; this is a bounded stress example, not an exhaustive maximum over every chart. BPHS/PVR emit 88 pair summaries; Rath emits 132. Chara always returns twelve rows and at most twelve intervals. The request cannot introduce an unbounded body/reference product. No global caches, output pruning or tolerance relaxation were introduced. These machine-specific measurements are cost evidence, not source or predictive validation.

The full engine docstring preflight passes 370 Python files. Scoped Ruff passes the sixteen other owned Python files. The two broad export owners retain their pre-existing lint debt: moira/__init__.py has 60 E402 findings and moira/facade.py has five F811 findings, matching their Git-baseline counts; no global lint-clean claim is made. Python 3.10 grammar admission passes all eighteen owned new/changed Python files. Documentation consistency, unchanged release identity 6.9.9, generated Hellenistic inventory, generated REST reference and wiki mirror guards pass:

.\.venv\Scripts\ruff.exe check moira/_jaimini_profile_inputs.py moira/jaimini_profile_policy.py moira/chara_profiles.py moira/argala_assessment.py moira/_facade_vedic.py moira/jaimini_extended.py moira/vedic.py moira_server/models/jaimini_profiles.py moira_server/services/jaimini_profiles.py moira_server/models/vedic_extended.py moira_server/routers/vedic_extended.py tests/unit/test_chara_profiles.py tests/unit/test_argala_assessment.py tests/unit/test_jaimini_profile_admission.py tests/server/test_server_jaimini_profiles.py tests/tools/benchmark_jaimini_profiles.py --no-fix
.\.venv\Scripts\python.exe tests/unit/test_docstring_governance.py
.\.venv\Scripts\python.exe scripts/check_doc_consistency.py
.\.venv\Scripts\python.exe scripts/check_release_identity.py
.\.venv\Scripts\python.exe scripts/generate_hellenistic_inventory.py --check
.\.venv\Scripts\python.exe scripts/sync_rest_api_reference.py --check
.\.venv\Scripts\python.exe scripts/sync_git_wiki.py --check
git diff --check

The canonical authoring tree is wiki/. The four new canonical pages were first registered with scoped Git intent-to-add for tracked-file discovery, then included in the authorized publication manifest below. The generated mirror is written only by scripts/sync_git_wiki.py --no-prune. No unrelated light-box files are registered, edited or removed. Generated Hellenistic owners were refreshed and checked; their final contents do not differ from the baseline.

Full BPHS Chara, PVR 2014 Chara, Rath Narayana, Rangacharya Yogada and general Argala strength comparison remain the five discovery-only candidates, each with a concrete source/algorithm/fixture gate. They do not weaken the selected six-profile closure and are never accepted as executable selectors. No missing strength, unavailable quarter composition or source conflict is silently filled.

The user subsequently requested commit/push followed by an adversarial review. The publication manifest includes only VED-013 engine/public/REST code, independent fixtures, tests, benchmark evidence, canonical documentation and its generated mirror. Concurrent light-box and natal-discovery work is excluded. The complete 847-test selected acceptance slice passed as one refreshed run, with zero failures/errors/skips, strict known-issue handling and downloads disabled. The command is the combined feature/regression invocation above with --junitxml=C:\dev\outputs\ved013-implementation-2026-10-12\publication.xml. All 29 DE441 resource receipts executed. Full engine docstrings, documentation consistency, release identity, Hellenistic inventory and wiki synchronization also pass.

During publication, a concurrent edit to moira_server/app.py registered three natal-discovery routes. The whole-worktree REST inventory check therefore became stale. That unrelated file and its routes are excluded from this commit. The REST guard passes against the exact staged server snapshot, exported with git ls-files -- moira_server and git show :<path> into C:\dev\outputs\ved013-publication-2026-10-12\server-index, imported ahead of the working tree, and checked with the unchanged scripts/sync_rest_api_reference.py --check. No route was deleted, filtered or silenced inside the generator; the published candidate has 523 paths.

The generated wiki is committed and pushed first. The parent engine commit containing this checkpoint records the exact wiki gitlink. Version remains 6.9.9; source publication is separate from release, deployment and website adoption. The requested subsequent adversarial review is a distinct evidence pass, not preemptively claimed clean by this publication checkpoint.

The review of published engine 573570b51de85ebd012eb246b93eafee57114caf found three result-integrity/error-boundary defects. The user authorized their repair. All three are repaired locally; this checkpoint does not publish the new changes or reopen the five source-admission candidates.

1. moira/chara_profiles.py now binds seed/direction and all twelve sequence positions to the recipe and Lagna. Count direction, inclusive count, base, exception and dignity evidence are derived from the same owned source-rule helpers for computation and result validation. Reordered but arithmetically consistent rows and false derivation labels are rejected.
2. moira/argala_assessment.py now binds per-body character and reasons, group character, exact within-sign degrees, effect/reason/comparison state, special-third evidence and Rath's two side-specific factors to the actual chart/profile. Other profiles cannot carry even one Rath factor. The original forged unobstructed verdict and Venus 31-degree/4-degree witness both fail. Missing source strength remains indeterminate.
3. moira/_jaimini_profile_results.py provides a confined post-admission JaiminiInvariantError (RuntimeError) boundary. Construction failures (ValueError, TypeError, KeyError) now become HTTP 500 even before serialization. The Jaimini service uses the same failure type for corrupt results and response contracts. Request admission remains outside that boundary. A nonrepresentable caller epoch still raises ValueError/422, after valid source arithmetic is established. Manual invalid vessel construction continues to raise ValueError. No global exception handler, route, response schema, numerical policy or default changed.

New negative tests in tests/unit/test_jaimini_profile_result_invariants.py substitute evidence that is valid in isolation but contradicts the result's chart or profile. Server regressions cover seven result-construction locations, three exception classes, and four corrupted-result transports. Internal error details are not exposed. Existing hostile input tests still assert 422.

Project Python 3.14.3; downloads disabled, network-excluding selection, strict known issues (empty register). From C:\dev\moira:

$env:MOIRA_NO_DOWNLOAD='1'
$env:MOIRA_TEST_MODE='1'
$env:MOIRA_STRICT_KNOWN_ISSUES='1'
.\.venv\Scripts\python.exe -m pytest tests/unit/test_jaimini_profile_result_invariants.py tests/unit/test_chara_profiles.py tests/unit/test_argala_assessment.py tests/unit/test_jaimini_profile_admission.py tests/server/test_server_jaimini_profiles.py tests/unit/test_jaimini_extended.py tests/unit/test_jaimini.py tests/unit/test_vedic_chara_d60_admission.py tests/server/test_vedic_chara_d60_admission.py tests/server/test_vedic_direct_input_hardening.py tests/server/test_server_jaimini_service.py tests/server/test_server_jaimini_routes.py tests/unit/test_api_surface_adversarial_audit.py -m 'not external_network' -q --junitxml=C:/dev/outputs/ved013-repair-2026-10-12/acceptance.xml

951 tests passed, zero failures/errors/skips, including the real unpatched application lifecycle and 29 executed DE441 resource checks. This is the prior 847-test selection plus 104 new cases, not a full-engine certification. The new independent sign-walk sweep checks 3,024 charts / 36,288 duration rows; the count/quarter sweep checks 2,640 Argala cases. These independently implement the admitted decisions and do not establish additional classical authority. Printed source fixtures, exact binary boundary tests and legacy regression comparisons also pass. All six source/compatibility fixture files are unchanged from the published commit; no expected answers or tolerances were regenerated.

The initial focused run passed 396 cases and failed 14 new assertions that incorrectly expected JSON from the server's generic HTTP 500 response. All fourteen already returned 500, not 422. The tests were corrected to check the actual status and absence of internal details, without changing the shared error envelope. The subsequent focused run passed 175 cases; the complete 951-test run above supersedes both. Raw XML remains under C:\dev\outputs\ved013-repair-2026-10-12.

The existing benchmark tool now hashes the new internal error-boundary owner. The new repair performance receipt preserves the prior benchmark artifact and uses the same six-profile workload, maximum 22 Argala references and thirty warm repetitions. Every repetition retains full engine/HTTP equality and unchanged response sizes.

.\.venv\Scripts\python.exe -m tests.tools.benchmark_jaimini_profiles --output tests/artifacts/benchmarks/jaimini_profiles_repair_2026-10-12/receipt.json

Stronger evidence validation has measurable cost, especially for Rath's 132 pair summaries; no unchanged-performance claim is made. These local timings are neither a service latency guarantee nor accuracy evidence. No evidence pruning, caching or tolerance relaxation was used to obtain them.

Full engine docstring preflight passes 371 files. Scoped Ruff and Python 3.10 AST grammar checks pass all seven owned Python files. Documentation and release identity guards pass (still 6.9.9). Hellenistic and REST inventories pass against the published server tree with only the repaired Jaimini service overlaid, using C:\dev\outputs\ved013-repair-2026-10-12\check_candidate.py. The exact snapshot and receipt are saved beside that script. This accounts for the concurrent natal-discovery additions in app.py/openapi.py without altering them or incorporating their routes into the VED-013 documentation.

.\.venv\Scripts\ruff.exe check moira/_jaimini_profile_results.py moira/chara_profiles.py moira/argala_assessment.py moira_server/services/jaimini_profiles.py tests/unit/test_jaimini_profile_result_invariants.py tests/server/test_server_jaimini_profiles.py tests/tools/benchmark_jaimini_profiles.py --no-fix
.\.venv\Scripts\python.exe tests/unit/test_docstring_governance.py
.\.venv\Scripts\python.exe scripts/check_doc_consistency.py
.\.venv\Scripts\python.exe scripts/check_release_identity.py
.\.venv\Scripts\python.exe C:/dev/outputs/ved013-repair-2026-10-12/check_candidate.py
.\.venv\Scripts\python.exe scripts/sync_git_wiki.py --no-prune
.\.venv\Scripts\python.exe scripts/sync_git_wiki.py --check
git diff --check

The standard, this receipt and the work register record the repair; their Git wiki mirror is generated locally. Concurrent natal-discovery/light-box files remain outside the repair. No commit, push, release, deployment or website work is part of this repair checkpoint.

The user subsequently authorized committing and pushing the three repairs. The seven recorded engine/REST source hashes still match the repair benchmark receipt, and the retained acceptance XML confirms 951 tests with zero failures/errors/skips. No engine or test implementation changed after that acceptance run; it was not needlessly repeated for publication.

Full 371-file engine docstring preflight, scoped seven-file Ruff, release identity, documentation and generated-wiki checks pass. The unchanged candidate-check script is rerun from C:\dev\outputs\ved013-repair-publication-2026-10-12\check_candidate.py; its private server snapshot excludes the unrelated natal-discovery edits and checks the Hellenistic and REST inventories without modifying the generator. Source/compatibility fixtures remain unchanged, and Python 3.10 grammar admission still passes the seven owned Python files.

The publication manifest is the two repaired evaluators, internal invariant error boundary, Jaimini serializer, new adversarial unit tests, expanded HTTP tests, benchmark tool/receipt, three canonical documents and generated wiki. The wiki is committed and pushed first; the containing engine commit records its exact gitlink. Concurrent natal-discovery and light-box files are excluded. This supersedes the preceding checkpoint's unpublished status. Version remains 6.9.9; no release, deployment, website work or new source profile is included.