Potential fix for code scanning alert no. 18: Deserialization of user-controlled data

[crazyrokr]

Potential fix for https://github.com/crazyrokr/tickonomics/security/code-scanning/18

Best fix: stop deserializing untrusted bytes with torch.load and switch to a safe, schema-validated format (JSON) for model state transport.

In this code, the safest minimal functional change is:

• In train_autoencoder (analytics/app/services/anomaly/anomaly_service.py), replace the current base64-encoded torch.save output with a JSON-serializable dict of tensor lists (state_dict converted to plain lists), then base64-encode the JSON string for transport compatibility.
• In detect_anomalies, replace torch.load(io.BytesIO(raw), ...) with:
  1. base64 decode,
  2. UTF-8 decode,
  3. json.loads,
  4. strict structure/type checks,
  5. rebuild tensors via torch.tensor(..., dtype=torch.float32).
• Keep existing layer-key checks and behavior unchanged otherwise.

This removes unsafe pickle deserialization while preserving endpoint contract (model_state remains a string) and anomaly detection behavior.

Suggested fixes powered by Copilot Autofix. Review carefully before merging.

[github-actions]

Benchmark Regression Report

Engine: same-runner A/B (base ea1ebf1 vs head 990e60e on a single job), gated by compare_benchmarks.py at >10% mean regression. See ADR-030.

benchmark	base mean (us)	head mean (us)	delta	status
test_fama_french_capm	145.400	144.939	-0.32%	ok

Runner: base = AMD EPYC 9V74 80-Core Processor; head = AMD EPYC 9V74 80-Core Processor; same runner = True.
Result: NO REGRESSION (0 of 1 benchmark(s) exceeded +10.00% on mean).