Harden MollyGraph smoke and relation cleanup

[brianmeyer]

Summary

• add fixture/no-download production smoke mode plus explicit real-model/download controls
• guard GLiNER2 loads with cache/download and memory checks, and keep heavy enrichment out of the default path
• suppress unsupported REPORTS_TO fan-out for simple employment statements with regression coverage
• update runtime docs/env examples for the memory-aware model flow

Verification

• service/.venv/bin/python -m pytest -q
• service/.venv/bin/python scripts/production_smoke.py --json
• git diff --check

Linear

Closes REC-183, REC-184, REC-185, REC-190, REC-191, REC-210.
REC-201 remains backlog for separate training-state hardening.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ff5ade1bf6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Enforce no-download mode for full model snapshot

When MOLLYGRAPH_ALLOW_MODEL_DOWNLOADS=0, this check treats the model as "offline-ready" if only model.safetensors is cached, but GLiNER2.from_pretrained(model_ref) is still called without any offline/local-only guard. On a clean machine with partial cache, GLiNER2 can still need other Hub files (e.g. config/tokenizer), so the runtime either performs network downloads despite the flag or fails unexpectedly in offline environments.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Make --real-models default truly disable downloads

Using env.setdefault("MOLLYGRAPH_ALLOW_MODEL_DOWNLOADS", "0") means --real-models does not actually enforce the default no-download behavior when the parent shell already exports MOLLYGRAPH_ALLOW_MODEL_DOWNLOADS=1. In that case, downloads stay enabled even without --allow-model-downloads, so the CLI flag semantics are inconsistent and can bypass the intended safety default.

Useful? React with 👍 / 👎.