v3.6.1

MemKraft 3.6.1

MemKraft 3.6.1 is the publishable patch release for the 3.6 Correction Learning line. The immutable v3.6.0 workflow passed its build and fresh-wheel gates but did not publish because PyPI Trusted Publishing was not configured.

Integrity and compatibility fixes

• Cryptographically binds the Improvement Ledger candidate and active artifact revision to the exact canonical correction payload injected into context. Mismatched or unrelated approved digests fail closed.
• Revalidates persisted visible correction text lengths and SHA-256 metadata during append-only replay, rejecting tampering while preserving the documented truncated-prefix behavior.
• Preserves the legacy combined-turn Hermes extraction contract for callers that do not provide role-message evidence. Callers that provide messages continue to receive role-separated user and assistant evidence.
• Updates benchmark and regression fixtures to use the same canonical correction digest as production governance.

Verification

• Focused integrity and Hermes regression suite: 65 passed.
• Correction, Improvement Ledger, benchmark, and Hermes integration suite: 317 passed.
• The full suite and release lineage gate are required to pass again on the protected-main candidate before tagging.

This patch does not move or replace the immutable v3.6.0 tag.