[Inference/Web] Surface model readiness and effective config before decompile jobs

[agorevski]

Finding

The web service can start with no usable model, but the health/UI flow does not make readiness or effective model configuration obvious before a user submits work.

Evidence

• load_model() logs failures and sets decompiler = None without failing process startup: web/app.py:337-352.
• /api/health always returns status: ok plus only model_loaded and mock_mode: web/app.py:943-950.
• The frontend starts by polling GPU stats, not health/readiness: web/static/app.js:1122-1126.
• Model config is rendered only after a decompile result arrives: web/static/app.js:497-510, web/static/app.js:1038-1041.

Impact

Users can paste bytecode and wait through analysis before discovering Solidity generation is unavailable. Operators also cannot use /api/health as a reliable readiness check for inference.

Recommended fix

Add a readiness endpoint/payload with ready, model_loaded, model_path, load error, lookup DB availability/stats, bytecode limits, concurrency limit, and effective generation defaults. Show this in the UI before input and disable or clearly label decompile when only TAC/lookup mode is available.

Acceptance criteria

• /api/health distinguishes process liveness from inference readiness.
• Health includes model path/config or last model load error, lookup availability, and request limits.
• UI displays a startup readiness banner/card before the first decompile.
• If no model is loaded, the Decompile button explains expected TAC-only/partial behavior before submission.

[agorevski]

Resolved in commit e4a1530. Validation: python -m compileall -q train.py download_hf_contracts.py src web scripts tests; uv run pytest -q (521 passed).