MemKraft 3.4.1

MemKraft 3.4.1

MemKraft 3.4.1 consolidates the completed but previously unpublished execution and governance work onto one release baseline.

The unpublished 3.3 candidate is absorbed into 3.4.1; there is no separately shipped 3.3 release line.

Highlights

• MKEP/0 execution Preview: append-only execution facts, deterministic projection, leases/fences, evidence gates, assessments, typed handoffs, a closed 15-operation dispatcher, CLI transport, read-only MCP projection, and cross-language conformance fixtures.
• Evaluation Corpus v0: deterministic, append-only registration of evaluation cases, labels, and runs. Correlation identifiers remain inert and do not write to the Improvement Ledger.
• Continual Improvement Ledger Preview: proposal and revision lineage, caller-supplied evaluation receipts, explicit promotion, activation and rollback records, idempotency, and corrupt-log fail-closed behavior. MemKraft records decisions but never executes artifact bodies or treats a model claim as authority.
• Project State Contract: deterministic drift findings across source, runtime, installed distribution, remote-backed Git state, working-tree cleanliness, durable checkout location, and project snapshots.
• Single release lineage: release_manifest.json binds each shipped feature to proposal, revision, evaluation, promotion, commits, and source paths. CI uses full Git history and blocks stranded implementation commits or changed package modules that are not owned by a declared feature.

Quantitative execution gates

The checked-in scripts/benchmark_execution_gates.py uses the locked 40-goal corpus, 20 warmups, and 200 repetitions. The historical candidate measurement is stored at docs/bench/results/mkep-3.3.0-g11-g12-macos-arm64.json:

• G11 projection p95 ratio: 0.9750× for 10,000 versus 1,000 records per goal, below the blocking 12× threshold.
• G12 CLI state.read p95: 61.313 ms versus 73.186 ms for memkraft --version, below the blocking +400 ms threshold.

These numbers describe the named local machine and candidate corpus, not universal latency. The final release reruns the deterministic benchmark gate before publication.

Scope exclusions

• Delay Retrospective Ledger is not shipped because no implementation commit exists. Its branch label pointed to the Evaluation Corpus commit.
• Project Memory Compiler remains a planned derived subsystem and is not represented as implemented or released.
• GitNexus code is not incorporated. Its current PolyForm Noncommercial 1.0.0 license is incompatible with direct absorption into this MIT package.

Compatibility and rollback

No migration is required. The execution, evaluation, improvement, and project-state stores are additive and lazy-created under .memkraft/. MemKraft 3.2.x ignores these stores on rollback, and re-upgrading resumes their retained append-only records.

MKEP/0 remains single-host and local-filesystem-only. Its gates and assessments are advisory; authority identities and artifact assertions are not authenticated. It provides no scheduler, graph engine, multi-host consensus, network-filesystem guarantee, or authorization mechanism.

Recovery release note

The immutable v3.4.0 tag failed before artifact publication because the release runner passed a relative PYTHONPATH=src into subprocess tests that change working directories. No PyPI distribution or GitHub Release was created from that tag. Version 3.4.1 uses the absolute GitHub workspace source path and adds a regression contract for that verifier boundary.