feat(ml): model trained on anonymized AI-usage dataset (discovery)

[Ar1anit]

Summary

Follow-up to #522 (anonymous per-feature token-usage tracking) and the dataset export (#TBD). Train an ML model on the anonymized usage dataset to derive insights about how users interact with AI features — the original goal behind tracking tokens.

Depends on #522 + the dataset export issue. This is research/discovery, scope to be refined once real data has accumulated.

Motivation

• Understand usage patterns: which feature sequences correlate with retention/upgrade, token-cost forecasting per cohort, anomaly/abuse detection.
• Produce insights that increase Smart Apply's valuation in a future sale.

Possible directions (to refine after data collection)

• Token-cost forecasting per feature / tier (time-series).
• Per-actorHash sequence modeling (which features get used together / in what order).
• Tier-upgrade propensity from usage patterns.
• Anomaly detection for abnormal token consumption (abuse / cost spikes).

Constraints

• Train only on the anonymized export — never on raw user records.
• No prompt/response content is available by design (feat(llm): anonymous per-feature token-usage tracking #522), so the model works on metadata only.

Acceptance criteria (initial)

• A documented hypothesis + chosen modeling approach.
• A baseline model + evaluation on the anonymized dataset.
• Written summary of insights suitable for a due-diligence deck.

Note: keep this issue open as an umbrella until enough data exists to make modeling worthwhile.