Academic Research Kernel 2.0.1

Academic Research Kernel 2.0.1 publishes the statistical-precision and concurrent-retraction fixes merged in PR #15, together with the release-path and tap-acceptance corrections in PR #16.

• Scientific notation, trailing zeros and explicit rounding boundaries now preserve reported p-value precision. Invalid/non-finite domains and incomplete statistical requests produce explicit failures; valid scientific mismatches remain normal results.
• Retraction runs commit DOI deltas with ordered observations and confirmations. Unknown responses retain last-known evidence, and concurrent older writes cannot restore stale state.
• MCP ping, notifications, malformed input recovery and shutdown are tested through real stdio. The unchanged raw sdist contains all self-test inputs.
• Publication validates the exact version/tag, source identity, internal versions and frozen file hashes. Same-hash partial-upload recovery does not rebuild or replace files.
• The tap gate verifies real remote discovery, installation, source records, complete files and enabled skills. It no longer accepts an empty installation merely because the upstream CLI exits zero.

Package/plugin version is 2.0.1. The 13 skills, 12 public MCP tools, receipt/schema v1 and license version axes are preserved. Historical 2.0.0 tags, releases and package files remain untouched.

Source commit: 534afa25f27bfd8c53cd26bcdfbd0bd430c8f66a
Source tree: f4f22d6b9e1c6106ba55fbc9c8961e51a939fb67

PR #15 merged as cab1beeddf68d714f5f453f5c0531cf7aebb3123; PR #16 merged normally at this release source. The accepted preparation tree equals the final merge tree.

Preparation closure CI: 19 successful jobs and one intentional PR tap skip. Final-main CI: all 20 jobs successful, including the actual tap lifecycle. Eleven platform environments each passed 1239 tests with no skips; macOS Intel passed 1236 tests with three existing optional Torch skips. All twelve environments passed 40 executable fences. Fresh wheel/sdist and clean-cache uvx passed on Linux, macOS and Windows.

The final tap report proves four targeted production GitHubSource discoveries, four exact-source complete installations and four enabled skills. Ordinary all-tap CLI search is not included in that claim. Automatic code review did not complete because the connector reported exhausted review quota; no completed-review or approval claim is made.

The source-bound release manifest records its generation-time publication state. Final public-channel receipts are recorded separately without changing the tagged source.

The selected formal bytes passed unchanged-sdist full acceptance on Windows x86_64 / Python3.12.6 / uv0.12.17 with CPU Torch2.14.1+cpu: 1239 tests, zero skips; static QA, 40 executable fences, i18n and compilation; separate wheel/sdist installed doctor, 12-tool discovery, real stdio and clean shutdown; independent empty-cache local-wheel uvx. SOURCE_DATE_EPOCH=1791454888. These two archives are the sole bytes selected for GitHub and PyPI.

Annotated tag object: bdfe285dd9150e3229bac9468e011dc161466f3f.

SHA-256:

• wheel: 7a986aab026e69a9979b43c0569aa0d18b90d42446704400ef5e8de839e411df
• sdist: 3a1a4ae18bf884281f9677aeeb3e95e5d831791d36d90d650edb2c8002ee761b

Public-channel closeout:

• PyPI 2.0.1 exposes both selected files. Anonymous downloads match the frozen GitHub bytes, sizes and SHA-256. Official-index no-cache installation and no-cache uvx passed doctor, 12-tool discovery, real stdio and shutdown. The published sdist passed 129 focused correctness regressions with zero skips. Optimized Python accepted expected 2.0.1 and rejected expected 2.0.0 with exit 1.
• Official MCP Registry exact 2.0.1 and latest returned HTTP 200, active/latest, with complete metadata equal to the frozen server.json. Published/updated at 2026-10-08T10:45:40.721046Z. Its exact PyPI/uvx/mcp/stdio pointer was independently executed with a new empty cache and passed version, doctor, 12-tool and stdio checks.
• Initial upload run uploaded the two files once, then failed the strict post-upload clean-source gate. The pinned publisher creates an unignored .github/.tmp generated action. Verify-only recovery succeeded with upload_count=0 and the publisher skipped; no package was reuploaded and no source/tag/asset changed.
• The first Registry submission was rejected with an expired short-lived JWT (HTTP 401). The same existing GitHub identity was refreshed; the second submission succeeded after confirming the exact version was still absent. No new account or long-lived credential was introduced.
• Final old-2.0.0 checks found no tag, Release, five-asset, PyPI-byte or Registry-history changes. Only the derived old Registry isLatest flag changed to false after 2.0.1 became latest.

The source manifest's publication=pending values are its generation-time history. This final receipt describes completed public-channel acceptance without changing the tagged source. The first upload run remains recorded as a failure; the successful verify-only run and independent public execution provide the completed acceptance evidence.