v2.0.0

[2.0.0] - 2026-08-18

v2.0.0 — stable Shadow read path with a default-on external derived cache,
Strict Read Only graph protection, Markdown authority, and fail-closed
Markdown/BM25 fallback. Biological memory, Logseq DB Safe-Sync writes, and
proactive adaptive runtime remain outside this release's enabled behavior.

Added

• Frozen Shadow read-profile fixtures — publish a content-free v1 admission
  corpus with healthy, malformed, future-version, unhealthy, and foreign-binding
  packets for deterministic compatibility checks; no cache operation or new read
  capability is introduced.

• Content-free Shadow status read target — add read_graph_data(target_type="shadow_status") as a versioned, query-only status surface with no cache initialization, schema migration, mutation, or content read.

• Versioned Shadow read profile — expose an additive, content-free shadow_db.read_profile with a closed profile version, installed producer version, hashed graph binding, committed generation, readiness, schema compatibility, and bounded capabilities so external read-side consumers can fail closed without filesystem or metadata-table access.

• Provider-free graph-outcome contracts (#483) — add frozen, closed, byte-stable task, environment, episode, event, rubric, veto, report, and content-free receipt schemas for resettable Logseq outcome evaluation; malformed provenance, duplicate identifiers, impossible event chains, unbounded budgets, incompatible write policies, and incomplete evidence fail closed without filesystem, vault, Shadow, model, provider, or network side effects.

• Local BEAM input provenance adapter (#448) — normalize caller-supplied BEAM chat and probing-question JSON into immutable, digest-bound input evidence with an exact public dataset pin; no downloads, inference, vault access, scoring, or result writes.

• Public-suite local-input provenance (#448) — require LoCoMo and LongMemEval callers to bind exact repository/revision/license metadata and a raw-input SHA-256 before parsing; persist that digest in immutable datasets and comparative manifests so mismatched public inputs fail closed without downloading data or executing systems.

• Retained synthetic retrieval scorecard evidence (#448) — commit the first manifest-bound, retrieval-only Italian hard-negative result with its deterministic fingerprint, bootstrap intervals, signal ablation, source revision, and a regression check that separates reproducible retrieval evidence from host-dependent latency and RSS diagnostics.

• Governed comparative cohort receipts (#448) — assemble deterministic, privacy-safe evidence receipts only when an exact public corpus and every required opaque result artifact have matching retained attestations across the required Matryca and open-system controls; no benchmark execution, provider call, vault access, raw-content retention, or result write.

• Local LongMemEval-V2 input provenance adapter (#448) — validate caller-supplied question, trajectory, and mapping JSONL with explicit license/revision provenance and immutable input-only digests; no downloads, inference, vault access, scoring, or result writes.

• Local LoCoMo retrieval adapter (#448) — normalize an already acquired public LoCoMo dataset into deterministic session-turn and QA-evidence cases without downloads, image fetching, model calls, vault access, or result persistence.

• Local LongMemEval retrieval adapter (#448) — normalize acquired cleaned/oracle JSON into digest-pinned, order-preserving in-memory cases with separate session and answer-marked-turn evidence, without downloads, model calls, vault access, or result persistence.

• Cross-system benchmark contracts (#448) — add closed, content-free manifests and retained-artifact reports that pin public-suite revisions, models, budgets, runtime context, and comparable Matryca/external controls without collecting prompts, answers, vault content, or running remote services.

• Semantic clustering scorecard completeness (#448) — extend the deterministic five-seed benchmark with mathematically correct ARI, predicted/expected cluster counts, explicit collapse detection and largest-cluster fraction, plus reproducible bootstrap intervals for aggregate metrics.

• Versioned clustering scorecard artifacts (#448) — add deterministic collapse-rate diagnostics and source-commit/fixture metadata to retained --output artifacts while preserving compact historical stdout and refusing overwrite or dirty-tree artifacts.

• Local retrieval execution bridge (#448) — replay caller-supplied or synthetic input evidence through an explicitly configured provider-neutral seam, rejecting malformed result sequences before artifact creation and retaining deterministic item, exclusion, failure, and digest-bound artifacts without downloads, inference, vault access, or network calls.

• Evidence archive foundation — add an immutable, privacy-safe, externally stored and idempotent P0 evidence-event archive for future governed memory candidates.

• Governed P0 evidence packet (#447) — add a byte-stable, retrieval-only coordination contract that binds exact source revision, canonical recall, benchmark scorecard, and append-only archive references without storage, model, remote, or canonical-write side effects.

Changed

• Evidence-first agentic-memory programme — sequence v2.1+ delivery through governed evidence, deterministic canonical recall, reproducible benchmarks, human-governed curation, and derived-only Shadow state before typed decay, procedural memory, or opt-in proactivity.
• BM25 scorecard benchmark schema — evolve the manifest-driven synthetic Italian hard-negative scorecard mode (--manifest-path) to schema v3 with explicit update/superseded labels, update-only accuracy, abstention precision/recall/confusion, deterministic ordered evidence fingerprints, and manifest digest/environment signatures; no runtime behavior changes.

Added

• Gated canonical recall (#186) — add the provider-free, graph-immutable search_graph(method="recall") P0 envelope with generation-bound UUID/hash references, deterministic reusable fingerprints, bounded per-turn retrieval, and explicit disabled/unavailable states; it uses only a fresh query-only Shadow FTS cache and never falls back, rebuilds, writes, or calls a model.

Fixed

• Read-only alias concurrency — serialize concurrent external X-Ray alias-state writes until the in-process lock is available, preserving bounded lock failure for all other page writes.
• Gate B RC2 artifact binding — bind fresh dual-profile qualification to the exact installed 2.0.0rc2 public wheel instead of rejecting it through the historical RC1 package identity.