demand_forecasting: enable time dimension + change-tracking note in predictive READMEs

[pkouki]

Summary

Related fixes to the predictive templates, surfaced while running demand_forecasting.py end-to-end:

1. demand_forecasting.py — enable the time dimension. The script previously set has_time_column=False with a stale "SDK limitation" workaround comment, treating Sale.date as a plain feature. Enabling time properly requires three coordinated changes (otherwise you hit ValueError: expect arity=3 for training/validation sets with has_time_column=True):

   • Set time_col=[Sale.date] on the PropertyTransformer.
   • Make the train/val/test relationships arity-3, carrying the timestamp ({Sale} at {Any:timestamp} has {Any:value}) and binding TrainTable.date / ValTable.date / TestTable.date.
   • Set has_time_column=True on the GNN.
   • Stale comments updated to match.

2. Add a rai init follow-up note to the predictive template READMEs telling users to add the following to their raiconfig.yaml:

   data:
       ensure_change_tracking: true

   Applied to: demand_forecasting, fraud-detection, smoker_status_prediction, subscriber_retention, telco_network_recovery.

3. fraud-detection README — add experiment-schema setup DDL. Mirrors the demand_forecasting README: a note under Prerequisites > Access with the SQL to create FRAUD_DETECTION.EXPERIMENTS and grant permissions to the RAI Native App before running the local demo.

4. Remove transient infra troubleshooting entries. These described temporary server-side states rather than user-actionable template issues:

   • demand_forecasting — "worker is not ready to accept jobs - please retry the job later"
   • subscriber_retention — "Train job stays in QUEUED with no progress"

Notes

I reviewed all predictive scripts across these templates for the same time-dimension issue. demand_forecasting.py was the only script with the arity mismatch. The other temporal scripts (fraud_detection*, retail_planning*) already use arity-3 relationships correctly, and the non-temporal ones (subscriber_retention, telco_network_recovery, smoker_status_prediction*) correctly leave has_time_column=False (static node tasks with non-temporal splits).

Test plan

• python demand_forecasting.py runs end-to-end with has_time_column=True (no arity error)
• Validation RMSE is reasonable vs. the predict-the-mean baseline
• python fraud_detection_local.py runs after the new schema DDL
• README yaml/sql blocks render correctly (consistent list indentation)

🤖 Generated with Claude Code

[github-actions]

The docs preview for this pull request has been deployed to Vercel!

✅ Preview:	https://relationalai-docs-msiut8k70-relationalai.vercel.app/build/templates
🔍 Inspect:	https://vercel.com/relationalai/relationalai-docs/AASkpaoMj9VWybPdYo5a3Psxzohy