Add genuine model-backed acceptance and clarify mocked integration coverage

[cgfixit]

Review follow-up, 2026-09-12. Rechecked origin at 98f267697b74.

Evidence and impact

Python 3.12.14 editable installation and 121 unit tests passed (89% coverage). CI run 34657207356 passed the unit matrix; its integration job was skipped.

The job is labeled “Integration Tests (full BERT pipeline),” but tests/integration/test_e2e.py patches SentenceTransformer and AutoTokenizer with mocks. Enabling this job alone does not prove weights/tokenizer compatibility or actual embedding inference. Its comment mentions a schedule, while the job condition permits workflow_dispatch or a commit-message marker.

This is a verification/claim gap, not a reproduced real-model crash.

Proposed implementation

1. Keep the fast mocked pipeline tests and label them accurately. Separate model-backed checks with a clear opt-in marker/job.
2. Choose the shipped supported model and record its identifier/revision, tokenizer revision, Python, torch and sentence-transformers versions. Permit a pre-populated local cache; make downloads explicit and bounded. Use synthetic/public sample text only.
3. Run the real model loader and tokenizer without monkeypatches, call actual embedding/extraction paths, and check expected dimensions, finite values, nonempty meaningful output, correct offsets/provenance and repeat-run behavior. Test empty/oversized input according to the existing contract.
4. Add failure-path coverage for unavailable/corrupt weights and offline cache misses. Report an explicit failed or unavailable real-model check rather than quietly switching to a mock.
5. Keep a small hand-labeled extraction set separate from the installation smoke. Report precision/recall or a similarly defined task metric; record cold/warm latency and peak memory on a named machine. Do not treat successful inference as proof of semantic quality.
6. Align workflow comments, trigger conditions and documentation. A manual acceptance job is sufficient initially; do not silently impose model downloads on every contributor run.

Acceptance criteria

• Existing mocked suite remains green and accurately named.
• One bounded run uses real weights and tokenizer and records reproducible environment details.
• The run cannot pass if the real model never loads.
• Model availability, inference correctness and task quality are reported separately.
• No private input text or credentials appear in CI artifacts.

Implementation should start by refreshing origin and checking this finding against the new head. Preserve existing safety controls. This issue/comment records findings and proposed acceptance criteria; it does not claim the fixes or unperformed live checks are complete.