LocalHold v0.2.0

• Changed malformed typed LOCALHOLD_* environment overrides to fail startup
  and operator commands without echoing their values; hold doctor now reports
  configuration failure and skips dependent probes when effective
  configuration is unavailable.
• Changed PostgreSQL startup to validate the resolved managed schema before
  serving: automatic migration bootstraps or repairs compatible schemas,
  disabled migration rejects absent or partial schemas, and all modes reject
  incompatible or conflicting schemas.
• Changed PostgreSQL schema initialization and migration to run atomically
  under a serialized advisory lock, with a configurable bound on each lock wait
  and exact migration-ledger validation.
• Added hold ui, an interactive terminal browser with search, detail,
  metadata, history, edit, expiry, and delete workflows. Browsing remains
  read-only; write access is acquired only for confirmed mutations.
• Added optimistic concurrency for interactive memory edits and deletes across
  SQLite and PostgreSQL. A separate record revision detects metadata and
  lifecycle changes without making unchanged content appear freshly updated.
• Added supported WAL-safe SQLite backup and restore commands with validation,
  dry runs, cross-process coordination, retained recovery snapshots, and
  transactional rollback when replacement or post-write validation fails.
• Added hold doctor with stable text and JSON diagnostics for configuration,
  storage schemas, embedding profiles, vector consistency, reranker health,
  and repairable versus blocking conditions.
• Added hold config init, hold config validate, and hold config paths for
  safe operator setup, validation, and active-path discovery.
• Added model artifact operator commands and hold embeddings status [--json]
  with provider health, secret-free configured/stored profile identity, vector
  coverage, consistency diagnostics, stable states, and automation-friendly
  exit codes.
• Added explicit reranker execution-provider policy with auto, cpu, and
  cuda modes, truthful compiled/selected/active reporting, required-mode
  startup enforcement, and process-safe model cache downloads.
• Replaced the built-in reranker raw ONNX graph with a checksum-pinned fused
  FP32 artifact and added an opt-in fused FP16 profile for explicit CUDA use.
  FP16 is rejected with CPU or auto fallback because it trades ranking
  precision for lower weight storage and faster CUDA reranking.
• Expanded release packaging with a deterministic Linux x86_64 CUDA 12 archive
  alongside the existing Linux CPU and Windows preview archives, and hardened
  them with pinned runtime inputs, private library loading, manifests, notices,
  checksums, extracted-artifact smoke tests, and Ubuntu 22.04/glibc-floor
  enforcement.
• Added a protected real-GPU release gate that validates the exact CUDA archive
  for native dependency closure, clean-host startup, FP32/FP16 parity,
  performance, and resource limits before publication.
• Added the LocalHold brand identity and maintained design-system assets.
• Hardened schema diagnostics, PostgreSQL index checks, redaction boundaries,
  pgvector-backed PostgreSQL CI coverage, release smoke tests, and CUDA runtime
  error reporting; removed retired API artifacts and made runtime time testable
  without wall-clock sleeps.
• Compatibility: opening a SQLite v1 database upgrades it to schema v2, and
  restore upgrades a strictly validated v1 backup only on a private staging
  copy. Older binaries reject schema v2 databases. PostgreSQL startup adds the
  matching record-revision migration when automatic migration is enabled.

LocalHold MiniLM L6 reranker artifacts v1

LocalHold MiniLM L6 reranker artifacts v1

These artifacts are derived from cross-encoder/ms-marco-MiniLM-L6-v2 at
immutable Hugging Face revision
c5ee24cb16019beea0893ab7796b1df96625c6b8.

Source files:

• onnx/model.onnx SHA-256:
  5d3e70fd0c9ff14b9b5169a51e957b7a9c74897afd0a35ce4bd318150c1d4d4a
• tokenizer.json SHA-256:
  d241a60d5e8f04cc1b2b3e9ef7a4921b27bf526d9f6050ab90f9267a1f9e5c66

The models were generated with Python 3.13, ONNX Runtime GPU 1.22.0, and
ONNX 1.18.0. The FP32 command was:

python -m onnxruntime.transformers.optimizer \
  --input model.onnx \
  --output ms-marco-MiniLM-L6-v2-fused-fp32.onnx \
  --model_type bert --num_heads 12 --hidden_size 384 --opt_level 0

The FP16 command added --float16 --use_gpu and wrote
ms-marco-MiniLM-L6-v2-fused-fp16.onnx.

Both transformations were run twice from the same source. Each repeated output
was byte-identical to its counterpart. Both outputs pass ONNX 1.18 model
validation and ONNX Runtime 1.22 inference on the intended providers.

The FP16 artifact is intended only for explicit CUDA execution. It can improve
reranking latency and reduce model weight storage, but reduced numerical
precision may change the ordering of closely scored documents. More testing is
needed to quantify ranking-quality effects across representative workloads.

LocalHold v0.1.0-beta.3

• Corrected MCP initialize metadata to report LocalHold's name and package
  version instead of the underlying SDK's identity.

LocalHold v0.1.0-beta.2

• Restored CPU reranker builds on the Ubuntu 22.04 release baseline by pinning
  the ONNX Runtime integration to a glibc-compatible release.
• Added an Ubuntu 22.04 reranker-enabled source test to pull request CI.