[maziyarpanahi]

Description

Add actual local training and offline export/evaluation commands for a small
clinical NLI encoder, with source-disjoint public/synthetic data and
fail-closed candidate promotion. This advances the remaining checkpoint
delivery work; it does not claim publication or enable a default alias.

Type of Change

• New feature
• Test addition/improvement
• Documentation update

Changes Made

• Pin the ClinicalE5-Small encoder and non-fiction public three-way NLI data.
  Attribute the author-published biomedical corpus; retain binary negatives
  as partial labels, not invented contradictions. Reject overlapping groups
  and remove whole groups with conflicting duplicate labels before splits.
• Train on the owner's local Apple hardware with seeded, configuration-backed
  AdamW; select checkpoints on validation only. No paid job, restricted data,
  hosted inference, model upload or visibility-changing operation.
• Export the same BERT sequence pooler/classifier to PyTorch, dynamic-axis
  ONNX int8 and MLX float32. Add the MLX offline runtime and inference-only
  torch context without introducing mandatory core dependencies.
• Evaluate the serving int8 export on final heldouts and the separate negation
  challenge. Preserve failed accuracy/calibration/coverage/parity gates in
  aggregate-only reports. Never accept all-abstention as successful coverage
  or fabricate a Hub revision from a candidate weight hash.

Testing

• Canonical make format, make lint, make format-check passed.
• Final full suite at ee5104a: 23,170 passed, 179 skipped, 20 existing warnings.
• Final focused training/NLI gate/backend tests: 59 passed.
• Native Apple torch/MLX padded segment-pair parity: 6 passed offline.
• Actual trained snapshot loaded/inferred through torch, ONNX int8 and
  MLX with sockets denied; actual dynamic ONNX/int8/MLX export preflight passed.
• Historical manually closed parent acceptance audit: 178 passed, 3 optional skips.
• Strict MkDocs build and wheel/sdist build passed; no package publishing.

Real candidate training/evaluation receipts will be added after the run finishes.
Synthetic unit tests are not model-quality evidence. Final holdouts are not
used for checkpoint selection.

Documentation

• Updated the clinical NLI model guide, data/license protocol, reproducible
  commands, optional training tools and quality/export boundaries.
• Added public helper docstrings.

Code Quality

• Scoped self-review and git diff --check.
• No unrelated formatting changes, downloaded text, weights, or PLANS files.

Dependencies

• No new mandatory package dependencies. Training and export use optional
  tools in an isolated environment; imports stay lazy at runtime boundaries.

Related Issues

Related to #3236 and #3233. Keep #3236 open until genuine candidate quality,
independent immutable model publication, offline resolution and registry
activation satisfy its checkpoint acceptance criteria.